diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index bbb967922..d9e468d7c 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -70,8 +70,8 @@ jobs:
     strategy:
       matrix:
         features:
-          - { label: "all", flags: "mmap,stopwords,lz4-compression,zstd-compression,failpoints,stemmer" }
-          - { label: "quickwit", flags: "mmap,quickwit,failpoints" }
+          - { label: "all", flags: "mmap,stopwords,lz4-compression,zstd-compression,failpoints,stemmer,jitexpr" }
+          - { label: "quickwit", flags: "mmap,quickwit,failpoints,jitexpr" }
           - { label: "none", flags: "" }
 
     name: test-${{ matrix.features.label}}
diff --git a/Cargo.toml b/Cargo.toml
index a4ff435ae..a5b050c78 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -76,6 +76,9 @@ measure_time = "0.9.0"
 arc-swap = "1.5.0"
 bon = "3.3.1"
 
+# EXPERIMENTAL. The API is likely to change in the near future.
+jitexpr = { version = "0.1", path = "./jitexpr", optional = true }
+
 columnar = { version = "0.7", path = "./columnar", package = "tantivy-columnar" }
 sstable = { version = "0.7", path = "./sstable", package = "tantivy-sstable", optional = true }
 stacker = { version = "0.7", path = "./stacker", package = "tantivy-stacker" }
@@ -84,7 +87,7 @@ tantivy-bitpacker = { version = "0.10", path = "./bitpacker" }
 common = { version = "0.11", path = "./common/", package = "tantivy-common" }
 tokenizer-api = { version = "0.7", path = "./tokenizer-api", package = "tantivy-tokenizer-api" }
 sketches-ddsketch = { version = "0.4", features = ["use_serde"] }
-datasketches = { version = "0.3.0", features = ["hll"] }
+datasketches = { version = "0.5.0", features = ["hll"] }
 futures-util = { version = "0.3.28", optional = true }
 futures-channel = { version = "0.3.28", optional = true }
 fnv = "1.0.7"
@@ -149,6 +152,7 @@ failpoints = ["fail", "fail/failpoints"]
 unstable = []                            # useful for benches.
 
 quickwit = ["sstable", "futures-util", "futures-channel"]
+jitexpr = ["dep:jitexpr"]
 
 # Compares only the hash of a string when indexing data.
 # Increases indexing speed, but may lead to extremely rare missing terms, when there's a hash collision.
diff --git a/benches/agg_bench.rs b/benches/agg_bench.rs
index 0ffd4322c..35a6d7096 100644
--- a/benches/agg_bench.rs
+++ b/benches/agg_bench.rs
@@ -11,13 +11,13 @@ use tantivy::aggregation::AggregationCollector;
 use tantivy::indexer::NoMergePolicy;
 use tantivy::query::{AllQuery, Query, TermQuery};
 use tantivy::schema::{IndexRecordOption, Schema, TextFieldIndexing, FAST, STRING};
-use tantivy::{doc, DateTime, Index, Term};
+use tantivy::{doc, DateTime, Index, Searcher, Term};
 
 #[global_allocator]
 pub static GLOBAL: &PeakMemAlloc<std::alloc::System> = &INSTRUMENTED_SYSTEM;
 
 type AggregationRequest = serde_json::Value;
-type AggregationExecutor = fn(&Index, AggregationRequest);
+type AggregationExecutor = fn(&Searcher, AggregationRequest);
 type BenchmarkConfig = (&'static str, AggregationRequest);
 type BenchmarkGroup = (&'static str, AggregationExecutor, Vec<BenchmarkConfig>);
 
@@ -40,6 +40,8 @@ fn main() {
 
     for (input_name, cardinality) in inputs {
         let index = get_test_index_bench(cardinality).unwrap();
+        let reader = index.reader().unwrap();
+        let searcher = reader.searcher();
         // On sparse this will not effectively filter anything. This should simulate co-located
         // data which are sparse.
         // So for sparse aggregation, although the value is sparse all values in the aggregation
@@ -51,26 +53,45 @@ fn main() {
             execute_agg_filtered
         };
         runner.set_name(input_name);
-        bench_agg(&mut runner, &index, execute_filtered);
+        bench_agg(&mut runner, &searcher, execute_filtered);
     }
 
-    bench_many_segments();
+    for num_segments in [100, 1_000] {
+        bench_many_segments(num_segments);
+    }
 }
 
-fn bench_many_segments() {
+fn bench_many_segments(num_segments: usize) {
     let mut runner = BenchRunner::new();
     runner.add_plugin(PeakMemAllocPlugin::new(GLOBAL));
-    let index = get_test_index_bench_with_num_segments(Cardinality::Full, 100).unwrap();
+    runner.config().set_num_iter_for_group(1);
+    let index = get_test_index_bench_with_num_segments(Cardinality::Full, num_segments).unwrap();
+    let reader = index.reader().unwrap();
+    let searcher = reader.searcher();
     let mut group = runner.new_group();
-    group.set_name("100_segments");
+    group.set_name(format!("{num_segments}_segments"));
+    let mut multi_terms_top500 = multi_terms_many_and_zipf_1000();
+    multi_terms_top500["mt"]["multi_terms"]["size"] = json!(500);
+    let mut nested_terms_top500 = nested_terms_many_and_zipf_1000();
+    // This limits outer buckets, unlike the global tuple limit for multi_terms.
+    nested_terms_top500["my_texts"]["terms"]["size"] = json!(500);
     for (benchmark_name, agg_req) in [
         ("terms_7", terms_on_field("text_few_terms_status")),
         ("terms_zipfs_1000", terms_on_field("text_1000_terms_zipf")),
         ("terms_150_000", terms_on_field("text_many_terms")),
         ("terms_all_unique", terms_on_field("text_all_unique_terms")),
+        benchmark_config!(nested_terms_status_and_zipf_1000),
+        benchmark_config!(multi_terms_status_and_zipf_1000),
+        benchmark_config!(nested_terms_many_and_zipf_1000),
+        benchmark_config!(multi_terms_many_and_zipf_1000),
+        (
+            "nested_terms_many_and_zipf_1000_top500",
+            nested_terms_top500,
+        ),
+        ("multi_terms_many_and_zipf_1000_top500", multi_terms_top500),
     ] {
-        group.register_with_input(benchmark_name, &index, move |index| {
-            execute_agg(index, agg_req.clone())
+        group.register_with_input(benchmark_name, &searcher, move |searcher| {
+            execute_agg(searcher, agg_req.clone())
         });
     }
     group.run();
@@ -82,7 +103,7 @@ fn terms_on_field(field: &str) -> AggregationRequest {
     })
 }
 
-fn bench_agg(runner: &mut BenchRunner, index: &Index, execute_filtered: AggregationExecutor) {
+fn bench_agg(runner: &mut BenchRunner, searcher: &Searcher, execute_filtered: AggregationExecutor) {
     let multi_terms_vs_nested = vec![
         benchmark_config!(nested_terms_status_and_zipf_1000),
         benchmark_config!(multi_terms_status_and_zipf_1000),
@@ -140,6 +161,7 @@ fn bench_agg(runner: &mut BenchRunner, index: &Index, execute_filtered: Aggregat
             benchmark_config!(terms_status_with_histogram),
             benchmark_config!(terms_zipf_1000_with_histogram),
             benchmark_config!(terms_status_with_date_histogram),
+            benchmark_config!(terms_status_with_date_histogram_26_bits),
             benchmark_config!(terms_status_with_date_histogram_single_bucket),
             benchmark_config!(terms_status_with_date_histogram_4_buckets),
             benchmark_config!(terms_status_with_date_histogram_8_buckets),
@@ -192,8 +214,8 @@ fn bench_agg(runner: &mut BenchRunner, index: &Index, execute_filtered: Aggregat
         let mut group = runner.new_group();
         group.set_name(group_name);
         for (benchmark_name, agg_req) in configs {
-            group.register_with_input(benchmark_name, index, move |index| {
-                execute(index, agg_req.clone())
+            group.register_with_input(benchmark_name, searcher, move |searcher| {
+                execute(searcher, agg_req.clone())
             });
         }
         group.run();
@@ -528,6 +550,17 @@ fn terms_status_with_date_histogram() -> AggregationRequest {
     })
 }
 
+fn terms_status_with_date_histogram_26_bits() -> AggregationRequest {
+    json!({
+        "my_texts": {
+            "terms": { "field": "text_few_terms_status" },
+            "aggs": {
+                "over_time": { "date_histogram": { "field": "timestamp_26_bits", "fixed_interval": "134h" } }
+            }
+        }
+    })
+}
+
 /// Same flattened terms × date_histogram, but with `hard_bounds`. The timestamps span 0..120h; the
 /// bounds drop only the first and last hour (ms: 1h=3_600_000, 119h=428_400_000), so almost every
 /// doc is in-bounds. This exercises the collector's hard-bounds path: `bounds.contains` runs per
@@ -835,34 +868,32 @@ fn multi_terms_status_and_zipf_1000_avg_sub_agg() -> AggregationRequest {
     })
 }
 
-fn execute_agg(index: &Index, agg_req: AggregationRequest) {
-    execute_agg_with_query(index, agg_req, &AllQuery);
+fn execute_agg(searcher: &Searcher, agg_req: AggregationRequest) {
+    execute_agg_with_query(searcher, agg_req, &AllQuery);
 }
 
-fn execute_agg_filtered(index: &Index, agg_req: AggregationRequest) {
-    let filter_field = index.schema().get_field("filter_field").unwrap();
+fn execute_agg_filtered(searcher: &Searcher, agg_req: AggregationRequest) {
+    let filter_field = searcher.schema().get_field("filter_field").unwrap();
     let filter_query = TermQuery::new(
         Term::from_field_text(filter_field, "a"),
         IndexRecordOption::Basic,
     );
-    execute_agg_with_query(index, agg_req, &filter_query);
+    execute_agg_with_query(searcher, agg_req, &filter_query);
 }
 
-fn execute_agg_filtered_on_single_term(index: &Index, agg_req: AggregationRequest) {
-    let filter_field = index.schema().get_field("single_term").unwrap();
+fn execute_agg_filtered_on_single_term(searcher: &Searcher, agg_req: AggregationRequest) {
+    let filter_field = searcher.schema().get_field("single_term").unwrap();
     let filter_query = TermQuery::new(
         Term::from_field_text(filter_field, "single_term"),
         IndexRecordOption::Basic,
     );
-    execute_agg_with_query(index, agg_req, &filter_query);
+    execute_agg_with_query(searcher, agg_req, &filter_query);
 }
 
-fn execute_agg_with_query(index: &Index, agg_req: AggregationRequest, query: &dyn Query) {
+fn execute_agg_with_query(searcher: &Searcher, agg_req: AggregationRequest, query: &dyn Query) {
     let agg_req: Aggregations = serde_json::from_value(agg_req).unwrap();
     let collector = get_collector(agg_req);
 
-    let reader = index.reader().unwrap();
-    let searcher = reader.searcher();
     black_box(searcher.search(query, &collector).unwrap());
 }
 
@@ -1058,6 +1089,7 @@ fn get_test_index_bench_with_num_segments(
     let score_field_f64 = schema_builder.add_f64_field("score_f64", score_fieldtype.clone());
     let score_field_i64 = schema_builder.add_i64_field("score_i64", score_fieldtype);
     let date_field = schema_builder.add_date_field("timestamp", FAST);
+    let date_26_bits_field = schema_builder.add_date_field("timestamp_26_bits", FAST);
     let schema = schema_builder.build();
 
     if reuse_index && std::path::Path::new(&index_dir).try_exists()? {
@@ -1110,6 +1142,7 @@ fn get_test_index_bench_with_num_segments(
     {
         let mut rng = StdRng::from_seed([1u8; 32]);
         let mut filter_rng = StdRng::from_seed([2u8; 32]);
+        let mut timestamp_26_bits_rng = StdRng::from_seed([3u8; 32]);
         let mut index_writer = index.writer_with_num_threads(1, 400_000_000)?;
         if num_segments > 1 {
             index_writer.set_merge_policy(Box::new(NoMergePolicy));
@@ -1184,6 +1217,7 @@ fn get_test_index_bench_with_num_segments(
         let _val_max = 1_000_000.0;
         const SPAN_MS: i64 = 120 * 3600 * 1000; // 120 hours in ms
         const NOISE_MS: i64 = 2 * 3600 * 1000; // ±2h noise
+        const MAX_26_BIT_TIMESTAMP_SECS: i64 = (1 << 26) - 1;
         for i in 0..doc_with_value {
             let val: f64 = rng.random_range(0.0..1_000_000.0);
             let json = if rng.random_bool(0.1) {
@@ -1195,6 +1229,14 @@ fn get_test_index_bench_with_num_segments(
             let base_ms = (i as i64 * SPAN_MS) / doc_with_value as i64;
             let noise_ms = rng.random_range(-NOISE_MS..NOISE_MS);
             let ts_ms = (base_ms + noise_ms).clamp(0, SPAN_MS);
+            // Force the endpoints and randomize the interior so the column uses a 26-bit packed
+            // representation rather than the blockwise-linear codec.
+            let ts_26_bits_secs = match i {
+                0 => 0,
+                1 => 1,
+                i if i + 1 == doc_with_value => MAX_26_BIT_TIMESTAMP_SECS,
+                _ => timestamp_26_bits_rng.random_range(0..=MAX_26_BIT_TIMESTAMP_SECS),
+            };
             add_document(doc!(
                 single_term => "single_term",
                 text_field => "cool",
@@ -1209,6 +1251,7 @@ fn get_test_index_bench_with_num_segments(
                 score_field_f64 => lg_norm.sample(&mut rng),
                 score_field_i64 => val as i64,
                 date_field => DateTime::from_timestamp_millis(ts_ms),
+                date_26_bits_field => DateTime::from_timestamp_secs(ts_26_bits_secs),
             ))?;
             if cardinality == Cardinality::OptionalSparse {
                 for _ in 0..20 {
diff --git a/benches/merge_segments.rs b/benches/merge_segments.rs
index 7e911f7ec..89e06fe84 100644
--- a/benches/merge_segments.rs
+++ b/benches/merge_segments.rs
@@ -10,7 +10,8 @@ use std::io::{self, Write};
 use std::path::{Path, PathBuf};
 use std::sync::{Arc, RwLock};
 
-use binggan::{black_box, BenchRunner};
+use binggan::plugins::PeakMemAllocPlugin;
+use binggan::{black_box, BenchRunner, PeakMemAlloc, INSTRUMENTED_SYSTEM};
 use rand::prelude::*;
 use rand::rngs::StdRng;
 use rand::SeedableRng;
@@ -20,8 +21,11 @@ use tantivy::directory::{
     WritePtr,
 };
 use tantivy::indexer::{merge_filtered_segments, NoMergePolicy};
-use tantivy::schema::{Schema, TEXT};
-use tantivy::{doc, HasLen, Index, IndexSettings, Segment};
+use tantivy::schema::{Schema, FAST, TEXT};
+use tantivy::{doc, HasLen, Index, IndexSettings, IndexSortByField, Order, Segment};
+
+#[global_allocator]
+static GLOBAL: &PeakMemAlloc<std::alloc::System> = &INSTRUMENTED_SYSTEM;
 
 #[derive(Clone, Default, Debug)]
 struct NullDirectory {
@@ -196,6 +200,91 @@ fn build_index(
     }
 }
 
+/// Like [`build_index`], sorted by a `u64` key that interleaves docs across segments.
+fn build_sorted_index(
+    num_segments: usize,
+    docs_per_segment: usize,
+    tokens_per_doc: usize,
+    vocab_size: usize,
+) -> MergeScenario {
+    let mut schema_builder = Schema::builder();
+    let body = schema_builder.add_text_field("body", TEXT);
+    let sort = schema_builder.add_u64_field("sort", FAST);
+    let schema = schema_builder.build();
+    let index = Index::builder()
+        .schema(schema)
+        .settings(IndexSettings {
+            sort_by_field: Some(IndexSortByField {
+                field: "sort".into(),
+                order: Order::Asc,
+            }),
+            ..Default::default()
+        })
+        .create_in_ram()
+        .unwrap();
+
+    assert!(vocab_size > 0);
+    let total_tokens = num_segments * docs_per_segment * tokens_per_doc;
+    let use_unique_terms = vocab_size >= total_tokens;
+    let mut rng = StdRng::from_seed([7u8; 32]);
+    let mut next_token_id: u64 = 0;
+
+    {
+        let mut writer = index.writer_with_num_threads(1, 256_000_000).unwrap();
+        writer.set_merge_policy(Box::new(NoMergePolicy));
+        for segment in 0..num_segments {
+            for row in 0..docs_per_segment {
+                let mut tokens = Vec::with_capacity(tokens_per_doc);
+                for _ in 0..tokens_per_doc {
+                    let token_id = if use_unique_terms {
+                        let id = next_token_id;
+                        next_token_id += 1;
+                        id
+                    } else {
+                        rng.random_range(0..vocab_size as u64)
+                    };
+                    tokens.push(format!("term_{token_id}"));
+                }
+                let sort_key = (row * num_segments + segment) as u64;
+                writer
+                    .add_document(doc!(body => tokens.join(" "), sort => sort_key))
+                    .unwrap();
+            }
+            writer.commit().unwrap();
+        }
+    }
+
+    let segments = index.searchable_segments().unwrap();
+    let settings = index.settings().clone();
+    let label = format!(
+        "segments={}, docs/seg={}, tokens/doc={}, vocab={}",
+        num_segments, docs_per_segment, tokens_per_doc, vocab_size
+    );
+
+    MergeScenario {
+        index,
+        segments,
+        settings,
+        label,
+    }
+}
+
+fn bench_merge(runner: &mut BenchRunner, group_name: String, scenario: MergeScenario) {
+    let mut group = runner.new_group();
+    group.set_name(group_name);
+    let segments = scenario.segments.clone();
+    let settings = scenario.settings.clone();
+    group.register("merge", move |_| {
+        let output_dir = NullDirectory::default();
+        let filter_doc_ids = vec![None; segments.len()];
+        let merged_index =
+            merge_filtered_segments(&segments, settings.clone(), filter_doc_ids, output_dir)
+                .unwrap();
+        black_box(merged_index);
+    });
+    group.run();
+}
+
 fn main() {
     let scenarios = vec![
         build_index(8, 50_000, 12, 8),
@@ -205,20 +294,21 @@ fn main() {
     ];
 
     let mut runner = BenchRunner::new();
+    runner.add_plugin(PeakMemAllocPlugin::new(GLOBAL));
     for scenario in scenarios {
-        let mut group = runner.new_group();
-        group.set_name(format!("merge_segments inv_index — {}", scenario.label));
-        let segments = scenario.segments.clone();
-        let settings = scenario.settings.clone();
-        group.register("merge", move |_| {
-            let output_dir = NullDirectory::default();
-            let filter_doc_ids = vec![None; segments.len()];
-            let merged_index =
-                merge_filtered_segments(&segments, settings.clone(), filter_doc_ids, output_dir)
-                    .unwrap();
-            black_box(merged_index);
-        });
+        let name = format!("merge_segments inv_index — {}", scenario.label);
+        bench_merge(&mut runner, name, scenario);
+    }
 
-        group.run();
+    let sorted_scenarios = vec![
+        build_sorted_index(8, 50_000, 12, 8),
+        build_sorted_index(16, 50_000, 12, 8),
+        build_sorted_index(16, 100_000, 12, 8),
+        build_sorted_index(8, 50_000, 8, 8 * 50_000 * 8),
+    ];
+
+    for scenario in sorted_scenarios {
+        let name = format!("merge_segments sorted inv_index — {}", scenario.label);
+        bench_merge(&mut runner, name, scenario);
     }
 }
diff --git a/bitpacker/src/bitpacker.rs b/bitpacker/src/bitpacker.rs
index 1efd2755e..dc192eeda 100644
--- a/bitpacker/src/bitpacker.rs
+++ b/bitpacker/src/bitpacker.rs
@@ -75,6 +75,7 @@ impl BitUnpacker {
     /// The bitunpacker works by doing an unaligned read of 8 bytes.
     /// For this reason, values of `num_bits` between
     /// [57..63] are forbidden.
+    #[inline]
     pub fn new(num_bits: u8) -> BitUnpacker {
         assert!(num_bits <= 7 * 8 || num_bits == 64);
         let mask: u64 = if num_bits == 64 {
@@ -101,7 +102,7 @@ impl BitUnpacker {
                 return 0;
             }
             let bit_shift = addr_in_bits & 7;
-            return self.get_slow_path(addr, bit_shift as u32, data);
+            return Self::get_slow_path(self.mask, addr, bit_shift as u32, data);
         }
         let bit_shift = addr_in_bits & 7;
         let bytes: [u8; 8] = (&data[addr..addr + 8]).try_into().unwrap();
@@ -110,8 +111,107 @@ impl BitUnpacker {
         val_shifted & self.mask
     }
 
+    /// Decodes consecutive values into `output`.
+    ///
+    /// Panics if the requested bits are outside `data`.
+    #[inline(always)]
+    pub fn get_range(&self, start_idx: u32, data: &[u8], output: &mut [u64]) {
+        if output.is_empty() {
+            return;
+        }
+        if self.num_bits == 0 {
+            output.fill(0);
+            return;
+        }
+        let start_idx = start_idx as usize;
+        let end_bit = (start_idx + output.len()) * self.num_bits;
+        debug_assert!(
+            end_bit.div_ceil(8) <= data.len(),
+            "Requested range is out of bounds"
+        );
+
+        // Fall back for ranges overlapping the end, where an eight-byte load would be partial.
+        let last_bit_addr = (start_idx + output.len() - 1) * self.num_bits;
+        // The optimization happening below requires reading full 8 bytes word.
+        // We check that data is long enough to allow for the last read, and if not, fall back for
+        // the following safe implementation
+        if (last_bit_addr >> 3) + 8 > data.len() {
+            for (offset, out) in output.iter_mut().enumerate() {
+                *out = self.get((start_idx + offset) as u32, data);
+            }
+            return;
+        }
+
+        let output_len = output.len();
+        /// # Safety
+        /// Eight bytes starting at `bit_addr >> 3` must fit in `data`.
+        #[inline(always)]
+        unsafe fn load(data: &[u8], bit_addr: usize) -> u64 {
+            // SAFETY: the caller guarantees that this load fits in `data`.
+            let packed = unsafe {
+                data.as_ptr()
+                    .add(bit_addr >> 3)
+                    .cast::<u64>()
+                    .read_unaligned()
+            };
+            u64::from_le(packed) >> (bit_addr & 7)
+        }
+        let mut bit_addr = start_idx * self.num_bits;
+        // Tantivy's `COLLECT_BLOCK_BUFFER_LEN` is 64, so optimize its common full-block case by
+        // decoding eight 1-8 bit values per load. Keep this literal in sync with that constant.
+        if output_len == 64 && self.num_bits <= 8 {
+            const VALUES_PER_CHUNK: usize = 8;
+            let (chunks, remainder) = output.as_chunks_mut::<VALUES_PER_CHUNK>();
+            debug_assert!(remainder.is_empty());
+            for chunk in chunks {
+                // SAFETY: the range-end check above guarantees that this load fits in `data`.
+                let packed: u64 = unsafe { load(data, bit_addr) };
+                for (i, out) in chunk.iter_mut().enumerate() {
+                    *out = (packed >> (i * self.num_bits)) & self.mask;
+                }
+                bit_addr += VALUES_PER_CHUNK * self.num_bits;
+            }
+            return;
+        }
+        const VALUES_PER_CHUNK: usize = 4;
+        let (chunks, remainder) = output.as_chunks_mut::<VALUES_PER_CHUNK>();
+        for chunk in chunks {
+            // Four values plus at most seven leading bits fit in one load.
+            // At 16 bits, values are byte-aligned, so there are no leading bits.
+            if self.num_bits <= 14 || self.num_bits == 16 {
+                // SAFETY: the range-end check above guarantees that this load fits in `data`.
+                let packed = unsafe { load(data, bit_addr) };
+                for (i, out) in chunk.iter_mut().enumerate() {
+                    *out = (packed >> (i * self.num_bits)) & self.mask;
+                }
+            } else if self.num_bits <= 28 || self.num_bits == 32 {
+                // Two values plus at most seven leading bits fit in one load.
+                // At 32 bits, values are byte-aligned, so there are no leading bits.
+                for (pair_idx, pair) in chunk.as_chunks_mut::<2>().0.iter_mut().enumerate() {
+                    // SAFETY: the range-end check above guarantees that this load fits in `data`.
+                    let packed = unsafe { load(data, bit_addr + pair_idx * 2 * self.num_bits) };
+                    pair[0] = packed & self.mask;
+                    pair[1] = (packed >> self.num_bits) & self.mask;
+                }
+            } else {
+                for (i, out) in chunk.iter_mut().enumerate() {
+                    // SAFETY: the range-end check above guarantees that this load fits in `data`.
+                    *out = unsafe { load(data, bit_addr + i * self.num_bits) } & self.mask;
+                }
+            }
+            bit_addr += VALUES_PER_CHUNK * self.num_bits;
+        }
+        for out in remainder {
+            // SAFETY: the range-end check above guarantees that this load fits in `data`.
+            *out = unsafe { load(data, bit_addr) } & self.mask;
+            bit_addr += self.num_bits;
+        }
+    }
+
+    // Pass the mask by value so specialized callers don't need to materialize a
+    // temporary BitUnpacker on the stack just to pass &self to this non-inlined helper.
     #[inline(never)]
-    fn get_slow_path(&self, addr: usize, bit_shift: u32, data: &[u8]) -> u64 {
+    fn get_slow_path(mask: u64, addr: usize, bit_shift: u32, data: &[u8]) -> u64 {
         let mut bytes: [u8; 8] = [0u8; 8];
         let available_bytes = data.len() - addr;
         // This function is meant to only be called if we did not have 8 bytes to load.
@@ -119,7 +219,7 @@ impl BitUnpacker {
         bytes[..available_bytes].copy_from_slice(&data[addr..]);
         let val_unshifted_unmasked: u64 = u64::from_le_bytes(bytes);
         let val_shifted = val_unshifted_unmasked >> bit_shift;
-        val_shifted & self.mask
+        val_shifted & mask
     }
 
     // Decodes the range of bitpacked `u32` values with idx
@@ -315,6 +415,33 @@ mod test {
             assert!(val <= max_val);
             assert_eq!(bitunpacker.get(i as u32, &buffer), val);
         }
+        for start in 0..=vals.len() {
+            let remaining = vals.len() - start;
+            for len in [0, remaining.min(1), remaining / 2, remaining] {
+                let mut output = vec![u64::MAX; len];
+                bitunpacker.get_range(start as u32, &buffer, &mut output);
+                assert_eq!(output, vals[start..start + len]);
+            }
+        }
+    }
+
+    #[test]
+    fn test_get_range_all_bit_widths() {
+        for num_bits in (0..=56).chain(std::iter::once(64)) {
+            let mask = u64::MAX.checked_shr(64 - num_bits as u32).unwrap_or(0);
+            for len in [0, 1, 2, 7, 8, 9, 31, 32, 33, 63, 64, 65, 255, 256, 257] {
+                let vals: Vec<u64> = (0..len)
+                    .map(|i| (i as u64).wrapping_mul(0x9e3779b97f4a7c15) & mask)
+                    .collect();
+                test_bitpacker_aux(num_bits, &vals);
+            }
+        }
+    }
+
+    #[test]
+    #[should_panic(expected = "Requested range is out of bounds")]
+    fn test_get_range_out_of_bounds() {
+        BitUnpacker::new(3).get_range(2, &[0], &mut [0]);
     }
 
     proptest::proptest! {
diff --git a/columnar/Cargo.toml b/columnar/Cargo.toml
index 6e07375c5..fb5e3912b 100644
--- a/columnar/Cargo.toml
+++ b/columnar/Cargo.toml
@@ -57,5 +57,9 @@ harness = false
 name = "bench_optional_index"
 harness = false
 
+[[bench]]
+name = "bench_multivalue_docids"
+harness = false
+
 [features]
 zstd-compression = ["sstable/zstd-compression"]
diff --git a/columnar/benches/bench_multivalue_docids.rs b/columnar/benches/bench_multivalue_docids.rs
new file mode 100644
index 000000000..e6b649feb
--- /dev/null
+++ b/columnar/benches/bench_multivalue_docids.rs
@@ -0,0 +1,68 @@
+use binggan::{InputGroup, black_box};
+use tantivy_columnar::{Column, ColumnarReader, ColumnarWriter};
+
+const NUM_DOCS: u32 = 1_000_000;
+const NUM_DISTINCT_VALUES: u64 = 1000;
+
+/// A multivalued column where `fill_percent` of the docs hold 1 to 4 values,
+/// spread over `NUM_DISTINCT_VALUES` distinct values.
+fn generate_multivalued_column(fill_percent: u32) -> Column {
+    let mut columnar_writer = ColumnarWriter::default();
+    for doc in 0..NUM_DOCS {
+        if doc % 100 >= fill_percent {
+            continue;
+        }
+        let num_values = 1 + doc % 4;
+        for value_idx in 0..num_values {
+            let value = (doc as u64 * 7 + value_idx as u64) % NUM_DISTINCT_VALUES;
+            columnar_writer.record_numerical(doc, "field", value);
+        }
+    }
+    let mut buffer: Vec<u8> = Vec::new();
+    columnar_writer
+        .serialize(NUM_DOCS, None, &mut buffer)
+        .unwrap();
+    let reader = ColumnarReader::open(buffer).unwrap();
+    reader.read_columns("field").unwrap()[0]
+        .open_u64_lenient()
+        .unwrap()
+        .unwrap()
+}
+
+fn main() {
+    let inputs: Vec<(String, Column)> = [100, 50, 10]
+        .into_iter()
+        .map(|fill_percent| {
+            (
+                format!("multi 1-4 values, {fill_percent}% docs"),
+                generate_multivalued_column(fill_percent),
+            )
+        })
+        .collect();
+    let mut group: InputGroup<Column> = InputGroup::new_with_inputs(inputs);
+
+    group.register("docids_all_values", |column: &Column| {
+        let mut doc_ids = Vec::new();
+        column.get_docids_for_value_range(0..=u64::MAX, 0..NUM_DOCS, &mut doc_ids);
+        black_box(doc_ids);
+    });
+    group.register("docids_1pct_values", |column: &Column| {
+        let mut doc_ids = Vec::new();
+        column.get_docids_for_value_range(0..=9, 0..NUM_DOCS, &mut doc_ids);
+        black_box(doc_ids);
+    });
+    // The block-wise fetch of a range query's doc set.
+    group.register("docids_all_values_blocks_of_1024", |column: &Column| {
+        let mut doc_ids = Vec::new();
+        let mut num_docs = 0;
+        for block_start in (0..NUM_DOCS).step_by(1024) {
+            let block_end = (block_start + 1024).min(NUM_DOCS);
+            doc_ids.clear();
+            column.get_docids_for_value_range(0..=u64::MAX, block_start..block_end, &mut doc_ids);
+            num_docs += doc_ids.len();
+        }
+        black_box(num_docs);
+    });
+
+    group.run();
+}
diff --git a/columnar/src/column/mod.rs b/columnar/src/column/mod.rs
index 3bc61cba0..5bf7963cb 100644
--- a/columnar/src/column/mod.rs
+++ b/columnar/src/column/mod.rs
@@ -47,10 +47,8 @@ impl<T: PartialOrd + Default> Column<T> {
 
 impl<T: MonotonicallyMappableToU64> Column<T> {
     pub fn to_u64_monotonic(self) -> Column<u64> {
-        let values = Arc::new(monotonic_map_column(
-            self.values,
-            StrictlyMonotonicMappingToInternal::<T>::new(),
-        ));
+        let values =
+            monotonic_map_column(self.values, StrictlyMonotonicMappingToInternal::<T>::new());
         Column {
             index: self.index,
             values,
diff --git a/columnar/src/column_index/multivalued_index.rs b/columnar/src/column_index/multivalued_index.rs
index ad7efd363..338f422c0 100644
--- a/columnar/src/column_index/multivalued_index.rs
+++ b/columnar/src/column_index/multivalued_index.rs
@@ -338,9 +338,7 @@ impl MultiValueIndexV2 {
         }
         ranks.truncate(write_doc_pos);
 
-        for rank in ranks.iter_mut() {
-            *rank = self.optional_index.select(*rank);
-        }
+        self.optional_index.select_batch(&mut ranks[..]);
     }
 }
 
diff --git a/columnar/src/column_values/mod.rs b/columnar/src/column_values/mod.rs
index 4911012af..87d8cca32 100644
--- a/columnar/src/column_values/mod.rs
+++ b/columnar/src/column_values/mod.rs
@@ -203,6 +203,11 @@ impl<T: Copy + PartialOrd + Debug + 'static> ColumnValues<T> for Arc<dyn ColumnV
         self.as_ref().get_val(idx)
     }
 
+    #[inline(always)]
+    fn get_vals(&self, indexes: &[u32], output: &mut [T]) {
+        self.as_ref().get_vals(indexes, output)
+    }
+
     #[inline(always)]
     fn get_vals_opt(&self, indexes: &[u32], output: &mut [Option<T>]) {
         self.as_ref().get_vals_opt(indexes, output)
diff --git a/columnar/src/column_values/monotonic_column.rs b/columnar/src/column_values/monotonic_column.rs
index 35de3787a..d5927982c 100644
--- a/columnar/src/column_values/monotonic_column.rs
+++ b/columnar/src/column_values/monotonic_column.rs
@@ -1,6 +1,8 @@
+use std::any::{Any, TypeId};
 use std::fmt::Debug;
 use std::marker::PhantomData;
 use std::ops::{Range, RangeInclusive};
+use std::sync::Arc;
 
 use crate::ColumnValues;
 use crate::column_values::monotonic_mapping::StrictlyMonotonicFn;
@@ -29,18 +31,26 @@ struct MonotonicMappingColumn<C, T, Input> {
 pub fn monotonic_map_column<C, T, Input, Output>(
     from_column: C,
     monotonic_mapping: T,
-) -> impl ColumnValues<Output>
+) -> Arc<dyn ColumnValues<Output>>
 where
     C: ColumnValues<Input> + 'static,
     T: StrictlyMonotonicFn<Input, Output> + Send + Sync + 'static,
     Input: PartialOrd + Debug + Send + Sync + Clone + 'static,
     Output: PartialOrd + Debug + Send + Sync + Clone + 'static,
 {
-    MonotonicMappingColumn {
+    // Preserve specialized codec methods (notably get_range) for identity mappings.
+    if T::IS_IDENTITY && TypeId::of::<Input>() == TypeId::of::<Output>() {
+        let column: Arc<dyn ColumnValues<Input>> = Arc::new(from_column);
+        return (&column as &dyn Any)
+            .downcast_ref::<Arc<dyn ColumnValues<Output>>>()
+            .unwrap()
+            .clone();
+    }
+    Arc::new(MonotonicMappingColumn {
         from_column,
         monotonic_mapping,
         _phantom: PhantomData,
-    }
+    })
 }
 
 impl<C, T, Input, Output> ColumnValues<Output> for MonotonicMappingColumn<C, T, Input>
@@ -104,6 +114,35 @@ mod tests {
         StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
     };
 
+    #[test]
+    fn test_u128_identity() {
+        let column: Arc<dyn ColumnValues<u128>> = monotonic_map_column(
+            VecColumn::from(vec![u128::MAX]),
+            StrictlyMonotonicMappingInverter::from(
+                StrictlyMonotonicMappingToInternal::<u128>::new(),
+            ),
+        );
+        assert_eq!(column.get_val(0), u128::MAX);
+    }
+
+    #[test]
+    fn test_same_type_non_identity() {
+        struct Shift;
+        impl StrictlyMonotonicFn<u64, u64> for Shift {
+            fn mapping(&self, value: u64) -> u64 {
+                value + 1
+            }
+
+            fn inverse(&self, value: u64) -> u64 {
+                value - 1
+            }
+        }
+        let column = monotonic_map_column(VecColumn::from(vec![1u64, 2, 3]), Shift);
+        let mut output = [0; 3];
+        column.get_range(0, &mut output);
+        assert_eq!(output, [2, 3, 4]);
+    }
+
     #[test]
     fn test_monotonic_mapping_iter() {
         let vals: Vec<u64> = (0..100u64).map(|el| el * 10).collect();
diff --git a/columnar/src/column_values/monotonic_mapping.rs b/columnar/src/column_values/monotonic_mapping.rs
index 4626053ed..f83f6d118 100644
--- a/columnar/src/column_values/monotonic_mapping.rs
+++ b/columnar/src/column_values/monotonic_mapping.rs
@@ -9,6 +9,9 @@ use crate::RowId;
 /// Monotonic maps a value to u64 value space.
 /// Monotonic mapping enables `PartialOrd` on u64 space without conversion to original space.
 pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Debug + Copy + Send + Sync {
+    /// Whether conversion to and from u64 leaves values unchanged.
+    const IS_IDENTITY: bool = false;
+
     /// Converts a value to u64.
     ///
     /// Internally all fast field values are encoded as u64.
@@ -32,6 +35,10 @@ pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Debug + Copy + Send
 /// so a value can be converted back to its original domain (e.g. ip address or f64) from its
 /// internal representation.
 pub trait StrictlyMonotonicFn<External, Internal> {
+    /// Whether both mapping directions leave values unchanged.
+    /// Only used to bypass mapping when the input and output types also match.
+    const IS_IDENTITY: bool = false;
+
     /// Strictly monotonically maps the value from External to Internal.
     fn mapping(&self, inp: External) -> Internal;
     /// Inverse of `mapping`. Maps the value from Internal to External.
@@ -58,6 +65,8 @@ impl<T> From<T> for StrictlyMonotonicMappingInverter<T> {
 impl<From, To, T> StrictlyMonotonicFn<To, From> for StrictlyMonotonicMappingInverter<T>
 where T: StrictlyMonotonicFn<From, To>
 {
+    const IS_IDENTITY: bool = T::IS_IDENTITY;
+
     #[inline(always)]
     fn mapping(&self, val: To) -> From {
         self.orig_mapping.inverse(val)
@@ -86,6 +95,8 @@ impl<External: MonotonicallyMappableToU128, T: MonotonicallyMappableToU128>
     StrictlyMonotonicFn<External, u128> for StrictlyMonotonicMappingToInternal<T>
 where T: MonotonicallyMappableToU128
 {
+    const IS_IDENTITY: bool = External::IS_IDENTITY;
+
     #[inline(always)]
     fn mapping(&self, inp: External) -> u128 {
         External::to_u128(inp)
@@ -101,6 +112,8 @@ impl<External: MonotonicallyMappableToU64, T: MonotonicallyMappableToU64>
     StrictlyMonotonicFn<External, u64> for StrictlyMonotonicMappingToInternal<T>
 where T: MonotonicallyMappableToU64
 {
+    const IS_IDENTITY: bool = External::IS_IDENTITY;
+
     #[inline(always)]
     fn mapping(&self, inp: External) -> u64 {
         External::to_u64(inp)
@@ -113,6 +126,8 @@ where T: MonotonicallyMappableToU64
 }
 
 impl MonotonicallyMappableToU64 for u64 {
+    const IS_IDENTITY: bool = true;
+
     #[inline(always)]
     fn to_u64(self) -> u64 {
         self
diff --git a/columnar/src/column_values/monotonic_mapping_u128.rs b/columnar/src/column_values/monotonic_mapping_u128.rs
index 9e16dc58c..5b6f5192e 100644
--- a/columnar/src/column_values/monotonic_mapping_u128.rs
+++ b/columnar/src/column_values/monotonic_mapping_u128.rs
@@ -4,6 +4,9 @@ use std::net::Ipv6Addr;
 /// Monotonic maps a value to u128 value space
 /// Monotonic mapping enables `PartialOrd` on u128 space without conversion to original space.
 pub trait MonotonicallyMappableToU128: 'static + PartialOrd + Copy + Debug + Send + Sync {
+    /// Whether conversion to and from u128 leaves values unchanged.
+    const IS_IDENTITY: bool = false;
+
     /// Converts a value to u128.
     ///
     /// Internally all fast field values are encoded as u64.
@@ -17,6 +20,8 @@ pub trait MonotonicallyMappableToU128: 'static + PartialOrd + Copy + Debug + Sen
 }
 
 impl MonotonicallyMappableToU128 for u128 {
+    const IS_IDENTITY: bool = true;
+
     fn to_u128(self) -> u128 {
         self
     }
diff --git a/columnar/src/column_values/u128_based/mod.rs b/columnar/src/column_values/u128_based/mod.rs
index d26f5ce35..b08026980 100644
--- a/columnar/src/column_values/u128_based/mod.rs
+++ b/columnar/src/column_values/u128_based/mod.rs
@@ -108,7 +108,7 @@ pub fn open_u128_mapped<T: MonotonicallyMappableToU128 + Debug>(
     let reader = CompactSpaceDecompressor::open(bytes)?;
     let inverted: StrictlyMonotonicMappingInverter<StrictlyMonotonicMappingToInternal<T>> =
         StrictlyMonotonicMappingToInternal::<T>::new().into();
-    Ok(Arc::new(monotonic_map_column(reader, inverted)))
+    Ok(monotonic_map_column(reader, inverted))
 }
 
 /// Returns the u64 representation of the u128 data.
diff --git a/columnar/src/column_values/u64_based/bitpacked.rs b/columnar/src/column_values/u64_based/bitpacked.rs
index 71319cbec..19083355b 100644
--- a/columnar/src/column_values/u64_based/bitpacked.rs
+++ b/columnar/src/column_values/u64_based/bitpacked.rs
@@ -1,18 +1,18 @@
 use std::io::{self, Write};
 use std::num::NonZeroU64;
 use std::ops::{Range, RangeInclusive};
+use std::sync::Arc;
 
 use common::{BinarySerializable, OwnedBytes};
 use fastdivide::DividerU64;
 use tantivy_bitpacker::{BitPacker, BitUnpacker, compute_num_bits};
 
 use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, ColumnStats};
-use crate::{ColumnValues, RowId};
+use crate::{ColumnValues, MonotonicallyMappableToU64, RowId};
 
-/// Depending on the field type, a different
-/// fast field is required.
+/// A bitpacked column reader. `u8::MAX` uses the bit width stored in the column.
 #[derive(Clone)]
-pub struct BitpackedReader {
+pub struct BitpackedReader<const NUM_BITS: u8 = { u8::MAX }> {
     data: OwnedBytes,
     bit_unpacker: BitUnpacker,
     stats: ColumnStats,
@@ -48,10 +48,37 @@ fn transform_range_before_linear_transformation(
     Some(start_before_gcd_multiplication..=end_before_gcd_multiplication)
 }
 
-impl ColumnValues for BitpackedReader {
+impl<const NUM_BITS: u8> BitpackedReader<NUM_BITS> {
+    #[inline(always)]
+    fn unpacker(&self) -> BitUnpacker {
+        if NUM_BITS == u8::MAX {
+            self.bit_unpacker
+        } else {
+            BitUnpacker::new(NUM_BITS)
+        }
+    }
+}
+
+impl<const NUM_BITS: u8> ColumnValues for BitpackedReader<NUM_BITS> {
     #[inline(always)]
     fn get_val(&self, doc: u32) -> u64 {
-        self.stats.min_value + self.stats.gcd.get() * self.bit_unpacker.get(doc, &self.data)
+        self.stats.min_value + self.stats.gcd.get() * self.unpacker().get(doc, &self.data)
+    }
+
+    fn get_range(&self, start: u64, output: &mut [u64]) {
+        debug_assert!(start <= u64::from(self.stats.num_rows));
+        debug_assert!(output.len() as u64 <= u64::from(self.stats.num_rows) - start);
+        if NUM_BITS == 0 {
+            output.fill(self.stats.min_value);
+            return;
+        }
+        self.unpacker().get_range(start as u32, &self.data, output);
+        let skip_processing = self.stats.gcd.get() == 1 && self.stats.min_value == 0;
+        if !skip_processing {
+            for val in output {
+                *val = self.stats.min_value + self.stats.gcd.get() * *val;
+            }
+        }
     }
     #[inline]
     fn min_value(&self) -> u64 {
@@ -139,11 +166,82 @@ impl ColumnCodec for BitpackedCodec {
     }
 }
 
+/// Specialize widths with a meaningful decoding speedup; other widths share one decoder.
+pub(super) fn load<T: MonotonicallyMappableToU64>(
+    bytes: OwnedBytes,
+) -> io::Result<Arc<dyn ColumnValues<T>>> {
+    let reader = BitpackedCodec::load(bytes)?;
+    macro_rules! specialize {
+        ($($bits:literal),* $(,)?) => {
+            match reader.bit_unpacker.bit_width() {
+                $(
+                    $bits => super::map_column_values::<_, T>(BitpackedReader::<$bits> {
+                        data: reader.data,
+                        bit_unpacker: reader.bit_unpacker,
+                        stats: reader.stats,
+                    }),
+                )*
+                _ => super::map_column_values::<_, T>(reader),
+            }
+        };
+    }
+    Ok(specialize!(1, 2, 3, 4, 5, 6, 7, 8, 16, 20, 24, 32, 64,))
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
     use crate::column_values::u64_based::tests::create_and_validate;
 
+    #[test]
+    fn test_specialized_bit_widths() {
+        for bits in (0..=56).chain(std::iter::once(64)) {
+            let mask = u64::MAX.checked_shr(64 - bits as u32).unwrap_or(0);
+            for gcd in [1, 3] {
+                if gcd != 1 && mask > (u64::MAX - 7) / gcd {
+                    continue;
+                }
+                let min_value = if bits == 64 { 0 } else { 7 };
+                let vals: Vec<u64> = (0..257)
+                    .map(|i| {
+                        let packed = match i {
+                            0 => 0,
+                            1 => mask.min(1),
+                            2 => mask,
+                            _ => (i as u64).wrapping_mul(0x9e3779b97f4a7c15) & mask,
+                        };
+                        min_value + gcd * packed
+                    })
+                    .collect();
+                let mut stats = super::super::StatsCollector::default();
+                for &val in &vals {
+                    stats.collect(val);
+                }
+                let mut buffer = Vec::new();
+                BitpackedCodecEstimator
+                    .serialize(&stats.stats(), &mut vals.iter().copied(), &mut buffer)
+                    .unwrap();
+                let data = OwnedBytes::new(buffer);
+                let reader = load::<u64>(data.clone()).unwrap();
+                let signed_reader = load::<i64>(data).unwrap();
+                for start in 0..=vals.len() {
+                    let mut output = vec![0; vals.len() - start];
+                    reader.get_range(start as u64, &mut output);
+                    assert_eq!(output, vals[start..]);
+                    for (i, &val) in output.iter().enumerate() {
+                        assert_eq!(reader.get_val((start + i) as u32), val);
+                    }
+                    let mut signed_output = vec![0; output.len()];
+                    signed_reader.get_range(start as u64, &mut signed_output);
+                    assert_eq!(
+                        signed_output,
+                        output.into_iter().map(i64::from_u64).collect::<Vec<_>>()
+                    );
+                }
+            }
+        }
+    }
+
     #[test]
     fn test_with_codec_data_sets_simple() {
         create_and_validate::<BitpackedCodec>(&[4, 3, 12], "name");
diff --git a/columnar/src/column_values/u64_based/mod.rs b/columnar/src/column_values/u64_based/mod.rs
index 42ef335ad..59ae64e09 100644
--- a/columnar/src/column_values/u64_based/mod.rs
+++ b/columnar/src/column_values/u64_based/mod.rs
@@ -114,7 +114,7 @@ impl CodecType {
         bytes: OwnedBytes,
     ) -> io::Result<Arc<dyn ColumnValues<T>>> {
         match self {
-            CodecType::Bitpacked => load_specific_codec::<BitpackedCodec, T>(bytes),
+            CodecType::Bitpacked => bitpacked::load::<T>(bytes),
             CodecType::Linear => load_specific_codec::<LinearCodec, T>(bytes),
             CodecType::BlockwiseLinear => load_specific_codec::<BlockwiseLinearCodec, T>(bytes),
         }
@@ -124,12 +124,16 @@ impl CodecType {
 fn load_specific_codec<C: ColumnCodec, T: MonotonicallyMappableToU64>(
     bytes: OwnedBytes,
 ) -> io::Result<Arc<dyn ColumnValues<T>>> {
-    let reader = C::load(bytes)?;
-    let reader_typed = monotonic_map_column(
+    Ok(map_column_values::<_, T>(C::load(bytes)?))
+}
+
+fn map_column_values<C: ColumnValues + 'static, T: MonotonicallyMappableToU64>(
+    reader: C,
+) -> Arc<dyn ColumnValues<T>> {
+    monotonic_map_column(
         reader,
         StrictlyMonotonicMappingInverter::from(StrictlyMonotonicMappingToInternal::<T>::new()),
-    );
-    Ok(Arc::new(reader_typed))
+    )
 }
 
 impl CodecType {
diff --git a/columnar/src/dynamic_column.rs b/columnar/src/dynamic_column.rs
index 58f689ebd..eadcbe736 100644
--- a/columnar/src/dynamic_column.rs
+++ b/columnar/src/dynamic_column.rs
@@ -1,5 +1,4 @@
 use std::net::Ipv6Addr;
-use std::sync::Arc;
 use std::{fmt, io};
 
 use common::file_slice::FileSlice;
@@ -124,11 +123,11 @@ impl DynamicColumn {
         match self {
             DynamicColumn::I64(column) => Some(DynamicColumn::F64(Column {
                 index: column.index,
-                values: Arc::new(monotonic_map_column(column.values, MapI64ToF64)),
+                values: monotonic_map_column(column.values, MapI64ToF64),
             })),
             DynamicColumn::U64(column) => Some(DynamicColumn::F64(Column {
                 index: column.index,
-                values: Arc::new(monotonic_map_column(column.values, MapU64ToF64)),
+                values: monotonic_map_column(column.values, MapU64ToF64),
             })),
             DynamicColumn::F64(_) => Some(self),
             _ => None,
@@ -142,7 +141,7 @@ impl DynamicColumn {
                 }
                 Some(DynamicColumn::I64(Column {
                     index: column.index,
-                    values: Arc::new(monotonic_map_column(column.values, MapU64ToI64)),
+                    values: monotonic_map_column(column.values, MapU64ToI64),
                 }))
             }
             DynamicColumn::I64(_) => Some(self),
@@ -157,7 +156,7 @@ impl DynamicColumn {
                 }
                 Some(DynamicColumn::U64(Column {
                     index: column.index,
-                    values: Arc::new(monotonic_map_column(column.values, MapI64ToU64)),
+                    values: monotonic_map_column(column.values, MapI64ToU64),
                 }))
             }
             DynamicColumn::U64(_) => Some(self),
diff --git a/jitexpr/Cargo.toml b/jitexpr/Cargo.toml
index e770cb802..85033708c 100644
--- a/jitexpr/Cargo.toml
+++ b/jitexpr/Cargo.toml
@@ -12,5 +12,9 @@ cranelift = "0.134.3"
 cranelift-jit = "0.134.3"
 cranelift-module = "0.134.3"
 cranelift-native = "0.134.3"
+lru = "0.18.2"
 regex = "1"
 thiserror = "2.0.1"
+
+[dev-dependencies]
+proptest = "1.7.0"
diff --git a/jitexpr/examples/basic.rs b/jitexpr/examples/basic.rs
index bc63bd7e1..d41dcc079 100644
--- a/jitexpr/examples/basic.rs
+++ b/jitexpr/examples/basic.rs
@@ -2,7 +2,7 @@ use std::collections::HashMap;
 use std::error::Error;
 use std::sync::Arc;
 
-use jitexpr::ast::{Function, InferredTypeSet, UntypedExpr, infer_types};
+use jitexpr::ast::{Function, InferredTypeSet, Literal, UntypedExpr, infer_types};
 use jitexpr::compile::{CompiledFn, CompiledFnCtx, compile};
 use jitexpr::types::{VarType, VariableValue};
 
@@ -13,7 +13,8 @@ fn main() -> Result<(), Box<dyn Error>> {
         Function::Add,
         vec![
             UntypedExpr::variable("my_col"),
-            UntypedExpr::literal(1.0f64),
+            // A float literal must be finite, so the conversion is fallible.
+            UntypedExpr::literal(Literal::try_from(1.0f64)?),
         ],
     )?;
 
diff --git a/jitexpr/src/ast/infer_types.rs b/jitexpr/src/ast/infer_types.rs
index 74b8191c3..9abe9bcb9 100644
--- a/jitexpr/src/ast/infer_types.rs
+++ b/jitexpr/src/ast/infer_types.rs
@@ -136,7 +136,7 @@ impl std::fmt::Display for InferredTypeSet {
     }
 }
 
-#[derive(Debug, thiserror::Error)]
+#[derive(Debug, thiserror::Error, Clone)]
 pub enum TypeError {
     #[error(transparent)]
     InvalidFnCall(#[from] InvalidFnCall),
diff --git a/jitexpr/src/ast/literal.rs b/jitexpr/src/ast/literal.rs
index 535290da8..4b3f8c6d6 100644
--- a/jitexpr/src/ast/literal.rs
+++ b/jitexpr/src/ast/literal.rs
@@ -1,6 +1,7 @@
 use std::sync::Arc;
 
 use crate::ast::InferredTypeSet;
+use crate::types::SafeF64;
 #[cfg(test)]
 use crate::types::VarType;
 
@@ -11,7 +12,7 @@ pub enum Literal {
     Bool(bool),
     U64(u64),
     I64(i64),
-    F64(f64),
+    F64(SafeF64),
     String(Arc<str>),
 }
 
@@ -43,10 +44,11 @@ impl Literal {
                 ..InferredTypeSet::NONE
             },
             Literal::F64(value) => {
+                let value: f64 = value.get();
                 let is_integral: bool = value.fract() == 0.0;
                 InferredTypeSet {
-                    i64: is_integral && *value >= i64::MIN as f64 && *value < -(i64::MIN as f64),
-                    u64: is_integral && *value >= 0.0 && *value < u64::MAX as f64,
+                    i64: is_integral && value >= i64::MIN as f64 && value < -(i64::MIN as f64),
+                    u64: is_integral && value >= 0.0 && value < u64::MAX as f64,
                     f64: true,
                     ..InferredTypeSet::NONE
                 }
@@ -86,9 +88,18 @@ impl From<i64> for Literal {
     }
 }
 
-impl From<f64> for Literal {
-    fn from(value: f64) -> Self {
-        Literal::F64(value)
+/// A float could not become a literal because it was NaN or infinite.
+#[derive(Clone, Copy, Debug, Eq, PartialEq, thiserror::Error)]
+#[error("an f64 literal must be finite, and neither NaN nor infinite")]
+pub struct NonFiniteFloat;
+
+impl TryFrom<f64> for Literal {
+    type Error = NonFiniteFloat;
+
+    /// Fails for NaN and for either infinity: [`SafeF64`] holds only finite
+    /// values, so those have no literal representation.
+    fn try_from(value: f64) -> Result<Literal, NonFiniteFloat> {
+        SafeF64::new(value).map(Literal::F64).ok_or(NonFiniteFloat)
     }
 }
 
@@ -108,6 +119,10 @@ impl From<&str> for Literal {
 mod tests {
     use super::*;
 
+    fn f64_literal(val: f64) -> Literal {
+        Literal::try_from(val).unwrap()
+    }
+
     #[test]
     fn test_literal_types_depend_on_representable_value() {
         let i64_f64 = InferredTypeSet {
@@ -125,8 +140,8 @@ mod tests {
         assert_eq!(Literal::I64(1).types(), InferredTypeSet::NUMERICAL);
         assert_eq!(Literal::I64(-1).types(), i64_f64);
         assert_eq!(Literal::U64(1 << 63).types(), u64_f64);
-        assert_eq!(Literal::F64(1.2).types(), InferredTypeSet::F64);
-        assert_eq!(Literal::F64(1.0).types(), InferredTypeSet::NUMERICAL);
+        assert_eq!(f64_literal(1.2).types(), InferredTypeSet::F64);
+        assert_eq!(f64_literal(1.0).types(), InferredTypeSet::NUMERICAL);
     }
 
     #[test]
@@ -162,22 +177,16 @@ mod tests {
     }
 
     #[test]
-    fn test_f64_literal_types_handle_integer_boundaries_and_special_values() {
+    fn test_f64_literal_types_handle_integer_boundaries() {
         assert_eq!(
-            Literal::F64(2f64.powi(63)).types(),
+            f64_literal(2f64.powi(63)).types(),
             InferredTypeSet {
                 u64: true,
                 f64: true,
                 ..InferredTypeSet::NONE
             }
         );
-        assert_eq!(Literal::F64(2f64.powi(64)).types(), InferredTypeSet::F64);
-        assert_eq!(Literal::F64(-0.0).types(), InferredTypeSet::NUMERICAL);
-        assert_eq!(Literal::F64(f64::NAN).types(), InferredTypeSet::F64);
-        assert_eq!(Literal::F64(f64::INFINITY).types(), InferredTypeSet::F64);
-        assert_eq!(
-            Literal::F64(f64::NEG_INFINITY).types(),
-            InferredTypeSet::F64
-        );
+        assert_eq!(f64_literal(2f64.powi(64)).types(), InferredTypeSet::F64);
+        assert_eq!(f64_literal(-0.0).types(), InferredTypeSet::NUMERICAL);
     }
 }
diff --git a/jitexpr/src/ast/mod.rs b/jitexpr/src/ast/mod.rs
index fe5028a54..de61743c1 100644
--- a/jitexpr/src/ast/mod.rs
+++ b/jitexpr/src/ast/mod.rs
@@ -1,11 +1,15 @@
 mod infer_types;
 mod literal;
+mod presence;
 mod serde;
 mod untyped_expr;
 
 pub use infer_types::{InferredTypeSet, TypeError, infer_types, infer_types_with_target};
 pub(crate) use infer_types::{infer_type_with_variable_types, infer_types_aux};
-pub use literal::Literal;
+pub use literal::{Literal, NonFiniteFloat};
+pub use presence::{
+    ConditionSet, VariablePresenceCondition, required_presence, required_presence_for_true,
+};
 pub(crate) use serde::format_variable_name;
 pub use serde::{DeserializeError, deserialize, serialize};
 pub use untyped_expr::UntypedExpr;
diff --git a/jitexpr/src/ast/presence.rs b/jitexpr/src/ast/presence.rs
new file mode 100644
index 000000000..edcf83fd0
--- /dev/null
+++ b/jitexpr/src/ast/presence.rs
@@ -0,0 +1,902 @@
+//! Necessary conditions on the presence of variables.
+//!
+//! Most functions return null as soon as one of their arguments is null. An expression can
+//! therefore often only produce a value, or only evaluate to `true`, if some of its variables are
+//! present. For instance, `(EQ (ADD a 1i64) b)` is null unless both `a` and `b` are present.
+//!
+//! A caller evaluating a predicate over many documents can use this to skip the documents missing
+//! these variables without evaluating the expression.
+
+use std::sync::Arc;
+
+use crate::ast::{Function, Literal, UntypedExpr};
+
+/// A boolean formula over the presence of variables.
+///
+/// It is meant to be used as a necessary condition: it is implied by some property of an
+/// expression (producing a value, or evaluating to `true`), but it does not imply it.
+///
+/// As much as possible, we try to normalize these object, in order to simplify them.
+#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub enum VariablePresenceCondition {
+    /// Always satisfied.
+    Always,
+    /// Never satisfied.
+    Never,
+    /// Satisfied when the variable is present.
+    Present(Arc<str>),
+    /// Satisfied when all of the conditions are satisfied.
+    All(ConditionSet),
+    /// Satisfied when at least one of the conditions is satisfied.
+    Any(ConditionSet),
+}
+
+/// The children of a [`VariablePresenceCondition::All`] or [`VariablePresenceCondition::Any`] node.
+///
+/// It can only be built through [`VariablePresenceCondition::all`] and
+/// [`VariablePresenceCondition::any`], which
+/// uphold the following hidden contract, on which the derived `Eq` and `Hash` rely:
+/// - children are sorted and distinct, and there are at least two of them;
+/// - no child is `Always`, `Never`, or a node of the same kind as the parent.
+#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct ConditionSet(Vec<VariablePresenceCondition>);
+
+impl ConditionSet {
+    /// Returns the children, in canonical order.
+    pub fn iter(&self) -> impl Iterator<Item = &VariablePresenceCondition> {
+        self.0.iter()
+    }
+}
+
+impl VariablePresenceCondition {
+    pub fn all(
+        conditions: impl IntoIterator<Item = VariablePresenceCondition>,
+    ) -> VariablePresenceCondition {
+        let mut children: Vec<VariablePresenceCondition> = Vec::new();
+        for condition in conditions {
+            match condition {
+                VariablePresenceCondition::Always => {}
+                VariablePresenceCondition::Never => return VariablePresenceCondition::Never,
+                VariablePresenceCondition::All(grand_children) => children.extend(grand_children.0),
+                condition => children.push(condition),
+            }
+        }
+        children.sort();
+        children.dedup();
+        match children.len() {
+            0 => VariablePresenceCondition::Always,
+            1 => children.pop().unwrap(),
+            _ => VariablePresenceCondition::All(ConditionSet(children)),
+        }
+    }
+
+    pub fn any(
+        conditions: impl IntoIterator<Item = VariablePresenceCondition>,
+    ) -> VariablePresenceCondition {
+        let mut children: Vec<VariablePresenceCondition> = Vec::new();
+        for condition in conditions {
+            match condition {
+                VariablePresenceCondition::Never => {}
+                VariablePresenceCondition::Always => return VariablePresenceCondition::Always,
+                VariablePresenceCondition::Any(grand_children) => children.extend(grand_children.0),
+                condition => children.push(condition),
+            }
+        }
+        children.sort();
+        children.dedup();
+        match children.len() {
+            0 => VariablePresenceCondition::Never,
+            1 => children.pop().unwrap(),
+            _ => VariablePresenceCondition::Any(ConditionSet(children)),
+        }
+    }
+
+    /// Evaluates the condition, given the presence of each variable.
+    #[cfg(test)]
+    pub fn eval(&self, is_present: &mut impl FnMut(&str) -> bool) -> bool {
+        match self {
+            VariablePresenceCondition::Always => true,
+            VariablePresenceCondition::Never => false,
+            VariablePresenceCondition::Present(variable_name) => is_present(variable_name),
+            VariablePresenceCondition::All(conditions) => conditions
+                .iter()
+                .all(|condition| condition.eval(&mut *is_present)),
+            VariablePresenceCondition::Any(conditions) => conditions
+                .iter()
+                .any(|condition| condition.eval(&mut *is_present)),
+        }
+    }
+}
+
+/// Returns a necessary presence condition for `expr` to evaluate to a non-null value.
+pub fn required_presence(expr: &UntypedExpr) -> VariablePresenceCondition {
+    match expr {
+        UntypedExpr::Literal(_) => VariablePresenceCondition::Always,
+        UntypedExpr::Variable(variable_name) => {
+            VariablePresenceCondition::Present(variable_name.clone())
+        }
+        UntypedExpr::FnCall { function, args } => required_presence_for_fn_call(*function, args),
+    }
+}
+
+/// Returns a necessary presence condition for `expr` to evaluate to a present `true`.
+pub fn required_presence_for_true(expr: &UntypedExpr) -> VariablePresenceCondition {
+    match expr {
+        UntypedExpr::Literal(Literal::Bool(true)) => VariablePresenceCondition::Always,
+        UntypedExpr::Literal(_) => VariablePresenceCondition::Never,
+        UntypedExpr::Variable(variable_name) => {
+            VariablePresenceCondition::Present(variable_name.clone())
+        }
+        UntypedExpr::FnCall { function, args } => {
+            required_presence_for_true_for_fn_call(*function, args)
+        }
+    }
+}
+
+fn required_presence_for_fn_call(
+    function: Function,
+    args: &[UntypedExpr],
+) -> VariablePresenceCondition {
+    // null argument as "null in, null out" would make callers skip matching documents.
+    match function {
+        // A null argument makes the result null.
+        //
+        // AND belongs here: `(AND false none)` is null.
+        Function::Abs
+        | Function::Add
+        | Function::And
+        | Function::Ceil
+        | Function::Concat
+        | Function::Divide
+        | Function::Eq
+        | Function::Floor
+        | Function::Gt
+        | Function::GtEq
+        | Function::IntMod
+        | Function::Left
+        | Function::Lower
+        | Function::Lt
+        | Function::LtEq
+        | Function::Max
+        | Function::Min
+        | Function::Multiply
+        | Function::Pow
+        | Function::RegexpExtract
+        | Function::Right
+        | Function::Round
+        | Function::SplitAfter
+        | Function::SplitBefore
+        | Function::Sqrt
+        | Function::Substring
+        | Function::SubstringCount
+        | Function::Subtract
+        | Function::TextJoin
+        | Function::Trim
+        | Function::Upper => VariablePresenceCondition::all(args.iter().map(required_presence)),
+        // OR is null only if all of its arguments are null.
+        Function::Or => VariablePresenceCondition::any(args.iter().map(required_presence)),
+        // IF is null if its condition is null. Otherwise it takes the presence of the selected
+        // branch.
+        Function::If => {
+            let [condition, when_true, when_false] = args else {
+                return VariablePresenceCondition::Always;
+            };
+            VariablePresenceCondition::all([
+                required_presence(condition),
+                VariablePresenceCondition::any([
+                    required_presence(when_true),
+                    required_presence(when_false),
+                ]),
+            ])
+        }
+        // These functions always return a present value.
+        //
+        // REGEXP_LIKE returns `false` for a null input.
+        Function::IsNotNull
+        | Function::IsNull
+        | Function::Neq
+        | Function::Not
+        | Function::RegexpLike => VariablePresenceCondition::Always,
+    }
+}
+
+fn required_presence_for_true_for_fn_call(
+    function: Function,
+    args: &[UntypedExpr],
+) -> VariablePresenceCondition {
+    match function {
+        Function::And => {
+            VariablePresenceCondition::all(args.iter().map(required_presence_for_true))
+        }
+        Function::Or => VariablePresenceCondition::any(args.iter().map(required_presence_for_true)),
+        Function::If => {
+            let [condition, when_true, when_false] = args else {
+                return VariablePresenceCondition::Always;
+            };
+            VariablePresenceCondition::all([
+                required_presence(condition),
+                VariablePresenceCondition::any([
+                    required_presence_for_true(when_true),
+                    required_presence_for_true(when_false),
+                ]),
+            ])
+        }
+        Function::IsNotNull => {
+            let [arg] = args else {
+                return VariablePresenceCondition::Always;
+            };
+            required_presence(arg)
+        }
+        // REGEXP_LIKE returns `false` for a null input.
+        Function::RegexpLike => {
+            let Some(input) = args.first() else {
+                return VariablePresenceCondition::Always;
+            };
+            required_presence(input)
+        }
+        // A `true` result is in particular a present result. This fallback is therefore correct
+        // for any function, including functions added later.
+        //
+        // It yields `Always` for NOT, NEQ, and IS_NULL, which are `true` when their
+        // argument is null.
+        _ => required_presence_for_fn_call(function, args),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+    use std::collections::hash_map::DefaultHasher;
+    use std::hash::{Hash, Hasher};
+
+    use proptest::prelude::*;
+    use proptest::strategy::BoxedStrategy;
+
+    use super::*;
+    use crate::ast::{InferredTypeSet, deserialize, infer_types_with_target};
+    use crate::compile::{StringArena, compile};
+    use crate::types::{VarType, VariableValue};
+
+    fn present(variable_name: &str) -> VariablePresenceCondition {
+        VariablePresenceCondition::Present(Arc::from(variable_name))
+    }
+
+    fn all(conditions: Vec<VariablePresenceCondition>) -> VariablePresenceCondition {
+        VariablePresenceCondition::all(conditions)
+    }
+
+    fn any(conditions: Vec<VariablePresenceCondition>) -> VariablePresenceCondition {
+        VariablePresenceCondition::any(conditions)
+    }
+
+    fn for_true(expr: &str) -> VariablePresenceCondition {
+        required_presence_for_true(&deserialize(expr).unwrap())
+    }
+
+    fn for_value(expr: &str) -> VariablePresenceCondition {
+        required_presence(&deserialize(expr).unwrap())
+    }
+
+    #[test]
+    fn test_all_simplification() {
+        assert_eq!(
+            VariablePresenceCondition::all([]),
+            VariablePresenceCondition::Always
+        );
+        assert_eq!(
+            VariablePresenceCondition::all([VariablePresenceCondition::Always, present("a")]),
+            present("a")
+        );
+        assert_eq!(
+            VariablePresenceCondition::all([present("a"), VariablePresenceCondition::Never]),
+            VariablePresenceCondition::Never
+        );
+        assert_eq!(
+            VariablePresenceCondition::all([
+                present("a"),
+                all(vec![present("b"), present("a")]),
+                present("c"),
+            ]),
+            all(vec![present("a"), present("b"), present("c")])
+        );
+        assert_eq!(
+            VariablePresenceCondition::all([any(vec![present("a"), present("b")]), present("c")]),
+            all(vec![any(vec![present("a"), present("b")]), present("c")])
+        );
+    }
+
+    #[test]
+    fn test_any_simplification() {
+        assert_eq!(
+            VariablePresenceCondition::any([]),
+            VariablePresenceCondition::Never
+        );
+        assert_eq!(
+            VariablePresenceCondition::any([VariablePresenceCondition::Never, present("a")]),
+            present("a")
+        );
+        assert_eq!(
+            VariablePresenceCondition::any([present("a"), VariablePresenceCondition::Always]),
+            VariablePresenceCondition::Always
+        );
+        assert_eq!(
+            VariablePresenceCondition::any([present("a"), any(vec![present("b"), present("a")])]),
+            any(vec![present("a"), present("b")])
+        );
+    }
+
+    fn hash_of(condition: &VariablePresenceCondition) -> u64 {
+        let mut hasher = DefaultHasher::new();
+        condition.hash(&mut hasher);
+        hasher.finish()
+    }
+
+    fn assert_same(left: VariablePresenceCondition, right: VariablePresenceCondition) {
+        assert_eq!(left, right);
+        assert_eq!(hash_of(&left), hash_of(&right));
+    }
+
+    #[test]
+    fn test_canonical_order() {
+        let (a, b, c) = (present("a"), present("b"), present("c"));
+        assert_same(
+            VariablePresenceCondition::all([a.clone(), b.clone()]),
+            VariablePresenceCondition::all([b.clone(), a.clone()]),
+        );
+        assert_same(
+            VariablePresenceCondition::any([c.clone(), a.clone(), b.clone()]),
+            VariablePresenceCondition::any([b.clone(), c.clone(), a.clone()]),
+        );
+        assert_ne!(
+            VariablePresenceCondition::all([a.clone(), b.clone()]),
+            VariablePresenceCondition::any([a.clone(), b.clone()])
+        );
+        let VariablePresenceCondition::All(children) =
+            VariablePresenceCondition::all([c.clone(), a.clone()])
+        else {
+            panic!("expected an All node");
+        };
+        assert_eq!(children.iter().collect::<Vec<_>>(), vec![&a, &c]);
+    }
+
+    #[test]
+    fn test_canonical_grouping_and_repetition() {
+        let (a, b, c) = (present("a"), present("b"), present("c"));
+        assert_same(
+            VariablePresenceCondition::all([
+                a.clone(),
+                VariablePresenceCondition::all([b.clone(), c.clone()]),
+            ]),
+            VariablePresenceCondition::all([
+                VariablePresenceCondition::all([c.clone(), a.clone()]),
+                b.clone(),
+            ]),
+        );
+        assert_same(
+            VariablePresenceCondition::all([a.clone(), a.clone()]),
+            a.clone(),
+        );
+        assert_same(
+            VariablePresenceCondition::any([
+                VariablePresenceCondition::all([a.clone(), b.clone()]),
+                VariablePresenceCondition::all([b.clone(), a.clone()]),
+            ]),
+            VariablePresenceCondition::all([a.clone(), b.clone()]),
+        );
+    }
+
+    /// A condition tree built without any normalization.
+    #[derive(Clone, Debug)]
+    enum RawCondition {
+        Always,
+        Never,
+        Present(usize),
+        All(Vec<RawCondition>),
+        Any(Vec<RawCondition>),
+    }
+
+    const RAW_VARIABLES: [&str; 4] = ["a", "b", "c", "d"];
+
+    impl RawCondition {
+        fn eval(&self, present_mask: u32) -> bool {
+            match self {
+                RawCondition::Always => true,
+                RawCondition::Never => false,
+                RawCondition::Present(variable_ord) => present_mask & (1 << variable_ord) != 0,
+                RawCondition::All(children) => {
+                    children.iter().all(|child| child.eval(present_mask))
+                }
+                RawCondition::Any(children) => {
+                    children.iter().any(|child| child.eval(present_mask))
+                }
+            }
+        }
+
+        /// Builds the canonical condition, visiting children in reverse order if `reverse`.
+        fn build(&self, reverse: bool) -> VariablePresenceCondition {
+            let build_children = |children: &[RawCondition]| {
+                let mut built: Vec<VariablePresenceCondition> =
+                    children.iter().map(|child| child.build(reverse)).collect();
+                if reverse {
+                    built.reverse();
+                }
+                built
+            };
+            match self {
+                RawCondition::Always => VariablePresenceCondition::Always,
+                RawCondition::Never => VariablePresenceCondition::Never,
+                RawCondition::Present(variable_ord) => present(RAW_VARIABLES[*variable_ord]),
+                RawCondition::All(children) => {
+                    VariablePresenceCondition::all(build_children(children))
+                }
+                RawCondition::Any(children) => {
+                    VariablePresenceCondition::any(build_children(children))
+                }
+            }
+        }
+    }
+
+    fn raw_conditions() -> impl Strategy<Value = RawCondition> {
+        let leaf = prop_oneof![
+            1 => Just(RawCondition::Always),
+            1 => Just(RawCondition::Never),
+            6 => (0..RAW_VARIABLES.len()).prop_map(RawCondition::Present),
+        ];
+        leaf.prop_recursive(4, 32, 4, |inner| {
+            prop_oneof![
+                prop::collection::vec(inner.clone(), 0..4).prop_map(RawCondition::All),
+                prop::collection::vec(inner, 0..4).prop_map(RawCondition::Any),
+            ]
+        })
+    }
+
+    /// Checks the hidden contract of `ConditionSet`, recursively.
+    fn assert_canonical(condition: &VariablePresenceCondition) {
+        let (children, is_all) = match condition {
+            VariablePresenceCondition::All(children) => (children, true),
+            VariablePresenceCondition::Any(children) => (children, false),
+            _ => return,
+        };
+        let children: Vec<&VariablePresenceCondition> = children.iter().collect();
+        assert!(children.len() >= 2, "{condition:?}");
+        assert!(
+            children.windows(2).all(|pair| pair[0] < pair[1]),
+            "{condition:?}"
+        );
+        for child in &children {
+            assert_canonical(child);
+            match (child, is_all) {
+                (VariablePresenceCondition::Always | VariablePresenceCondition::Never, _) => {
+                    panic!("neutral or absorbing child in {condition:?}")
+                }
+                (VariablePresenceCondition::All(_), true)
+                | (VariablePresenceCondition::Any(_), false) => {
+                    panic!("same-kind child in {condition:?}")
+                }
+                _ => {}
+            }
+        }
+    }
+
+    proptest! {
+        #[test]
+        fn proptest_canonical_form(raw in raw_conditions()) {
+            let condition = raw.build(false);
+            assert_canonical(&condition);
+            let reversed = raw.build(true);
+            prop_assert_eq!(&condition, &reversed);
+            prop_assert_eq!(hash_of(&condition), hash_of(&reversed));
+            for present_mask in 0..(1u32 << RAW_VARIABLES.len()) {
+                let mut is_present = |variable_name: &str| {
+                    let variable_ord =
+                        RAW_VARIABLES.iter().position(|name| *name == variable_name).unwrap();
+                    present_mask & (1 << variable_ord) != 0
+                };
+                prop_assert_eq!(condition.eval(&mut is_present), raw.eval(present_mask));
+            }
+        }
+    }
+
+    fn presence_of<'a>(present_names: &'a [&'a str]) -> impl FnMut(&str) -> bool + 'a {
+        move |variable_name: &str| present_names.contains(&variable_name)
+    }
+
+    #[test]
+    fn test_eval() {
+        let condition = all(vec![present("a"), any(vec![present("b"), present("c")])]);
+        assert!(condition.eval(&mut presence_of(&["a", "c"])));
+        assert!(!condition.eval(&mut presence_of(&["a"])));
+        assert!(!condition.eval(&mut presence_of(&["b", "c"])));
+        assert!(VariablePresenceCondition::Always.eval(&mut presence_of(&[])));
+        assert!(!VariablePresenceCondition::Never.eval(&mut presence_of(&["a"])));
+    }
+
+    #[test]
+    fn test_literals() {
+        assert_eq!(for_true("true"), VariablePresenceCondition::Always);
+        assert_eq!(for_true("false"), VariablePresenceCondition::Never);
+        assert_eq!(for_true("none"), VariablePresenceCondition::Never);
+        assert_eq!(for_value("none"), VariablePresenceCondition::Always);
+        assert_eq!(for_value("1u64"), VariablePresenceCondition::Always);
+        // `none` in a strict function is conservatively ignored.
+        assert_eq!(for_true("(EQ a none)"), present("a"));
+    }
+
+    #[test]
+    fn test_variable() {
+        assert_eq!(for_true("flag"), present("flag"));
+        assert_eq!(for_value("a"), present("a"));
+    }
+
+    #[test]
+    fn test_strict_functions() {
+        assert_eq!(
+            for_true("(EQ (ADD a 1i64) b)"),
+            all(vec![present("a"), present("b")])
+        );
+        assert_eq!(for_true("(GT (ABS a) 3i64)"), present("a"));
+        assert_eq!(
+            for_value(r#"(CONCAT "," "true" (UPPER a) (SUBSTRING b 0i64 2i64))"#),
+            all(vec![present("a"), present("b")])
+        );
+        assert_eq!(for_value(r#"(REGEXP_EXTRACT a "(x+)" 1u64)"#), present("a"));
+        assert_eq!(for_value("(ADD)"), VariablePresenceCondition::Always);
+    }
+
+    #[test]
+    fn test_and_or() {
+        assert_eq!(
+            for_true("(AND (EQ a 1i64) (LT b 2i64) c)"),
+            all(vec![present("a"), present("b"), present("c")])
+        );
+        assert_eq!(
+            for_true("(OR (EQ a 1i64) (EQ b 2i64))"),
+            any(vec![present("a"), present("b")])
+        );
+        assert_eq!(
+            for_true("(AND (OR (EQ a 1i64) (EQ b 2i64)) (EQ c 3i64))"),
+            all(vec![any(vec![present("a"), present("b")]), present("c")])
+        );
+        // AND is null as soon as one of its arguments is null.
+        assert_eq!(for_value("(AND (NOT a) b)"), present("b"));
+        assert_eq!(
+            for_value("(OR (EQ a 1i64) (EQ b 2i64))"),
+            any(vec![present("a"), present("b")])
+        );
+    }
+
+    #[test]
+    fn test_null_tolerant_functions() {
+        assert_eq!(
+            for_true("(NOT (EQ a 1i64))"),
+            VariablePresenceCondition::Always
+        );
+        assert_eq!(for_true("(NEQ a 1i64)"), VariablePresenceCondition::Always);
+        assert_eq!(for_true("(IS_NULL a)"), VariablePresenceCondition::Always);
+        assert_eq!(
+            for_value("(IS_NOT_NULL a)"),
+            VariablePresenceCondition::Always
+        );
+        assert_eq!(
+            for_true("(IS_NOT_NULL (ADD a b))"),
+            all(vec![present("a"), present("b")])
+        );
+        assert_eq!(
+            for_value(r#"(REGEXP_LIKE a "x")"#),
+            VariablePresenceCondition::Always
+        );
+        assert_eq!(for_true(r#"(REGEXP_LIKE a "x")"#), present("a"));
+        assert_eq!(
+            for_true("(OR (EQ a 1i64) (IS_NULL b))"),
+            VariablePresenceCondition::Always
+        );
+        assert_eq!(for_true("(AND (EQ a 1i64) (IS_NULL b))"), present("a"));
+    }
+
+    #[test]
+    fn test_if() {
+        assert_eq!(
+            for_value("(IF c a b)"),
+            all(vec![present("c"), any(vec![present("a"), present("b")])])
+        );
+        assert_eq!(
+            for_true("(IF c (EQ a 1i64) (EQ b 1i64))"),
+            all(vec![present("c"), any(vec![present("a"), present("b")])])
+        );
+        assert_eq!(for_true("(IF c true false)"), present("c"));
+        assert_eq!(
+            for_true("(IF c false false)"),
+            VariablePresenceCondition::Never
+        );
+        assert_eq!(for_value("(IF c 1i64 a)"), present("c"));
+    }
+
+    // The property tests below check that the conditions are indeed necessary, by comparing them
+    // with the compiled expression over random inputs.
+
+    const VARIABLES: [(&str, VarType); 7] = [
+        ("b0", VarType::Bool),
+        ("b1", VarType::Bool),
+        ("n0", VarType::I64),
+        ("n1", VarType::I64),
+        ("f0", VarType::F64),
+        ("s0", VarType::Str),
+        ("s1", VarType::Str),
+    ];
+
+    /// Values are indexed by variable, in the order of `VARIABLES`. `None` means null.
+    type Assignment = Vec<Option<u8>>;
+
+    fn variable_value(var_type: VarType, value_ord: u8) -> VariableValue<'static> {
+        let value_ord = value_ord as usize;
+        match var_type {
+            VarType::Bool => VariableValue::some([true, false, true, false][value_ord]),
+            VarType::I64 => VariableValue::some([0i64, 1, -2, 3][value_ord]),
+            VarType::F64 => VariableValue::some([0.0f64, 1.5, -1.0, 2.0][value_ord]),
+            VarType::Str => VariableValue::some(["", "a", "ab,a", "ba"][value_ord]),
+            VarType::U64 | VarType::None => unreachable!(),
+        }
+    }
+
+    fn assignments() -> impl Strategy<Value = Vec<Assignment>> {
+        let value = prop_oneof![Just(None), (0u8..4).prop_map(Some)];
+        prop::collection::vec(prop::collection::vec(value, VARIABLES.len()), 1..16)
+    }
+
+    struct ExprStrategies {
+        boolean: BoxedStrategy<String>,
+        number: BoxedStrategy<String>,
+        string: BoxedStrategy<String>,
+    }
+
+    fn leaves() -> ExprStrategies {
+        let pick = |choices: &'static [&'static str]| {
+            prop::sample::select(choices)
+                .prop_map(str::to_string)
+                .boxed()
+        };
+        ExprStrategies {
+            boolean: pick(&["b0", "b1", "true", "false", "none"]),
+            number: pick(&["n0", "n1", "f0", "0i64", "3i64", "-2i64", "1.5f64", "none"]),
+            string: pick(&["s0", "s1", r#""a""#, r#""""#, "none"]),
+        }
+    }
+
+    fn unary(arg: &BoxedStrategy<String>, template: &'static str) -> BoxedStrategy<String> {
+        arg.clone()
+            .prop_map(move |arg| template.replace("$0", &arg))
+            .boxed()
+    }
+
+    fn binary(
+        left: &BoxedStrategy<String>,
+        right: &BoxedStrategy<String>,
+        template: &'static str,
+    ) -> BoxedStrategy<String> {
+        (left.clone(), right.clone())
+            .prop_map(move |(left, right)| template.replace("$0", &left).replace("$1", &right))
+            .boxed()
+    }
+
+    fn ternary(
+        first: &BoxedStrategy<String>,
+        second: &BoxedStrategy<String>,
+        third: &BoxedStrategy<String>,
+        template: &'static str,
+    ) -> BoxedStrategy<String> {
+        (first.clone(), second.clone(), third.clone())
+            .prop_map(move |(first, second, third)| {
+                template
+                    .replace("$0", &first)
+                    .replace("$1", &second)
+                    .replace("$2", &third)
+            })
+            .boxed()
+    }
+
+    /// Returns strategies generating well-typed expressions of the given depth.
+    fn exprs(depth: u32) -> ExprStrategies {
+        let leaves = leaves();
+        if depth == 0 {
+            return leaves;
+        }
+        let ExprStrategies {
+            boolean: b,
+            number: n,
+            string: s,
+        } = exprs(depth - 1);
+        let any_kind = prop_oneof![b.clone(), n.clone(), s.clone()].boxed();
+        let boolean = prop::strategy::Union::new(vec![
+            leaves.boolean,
+            binary(&b, &b, "(AND $0 $1)"),
+            ternary(&b, &b, &b, "(AND $0 $1 $2)"),
+            binary(&b, &b, "(OR $0 $1)"),
+            ternary(&b, &b, &b, "(OR $0 $1 $2)"),
+            unary(&b, "(NOT $0)"),
+            unary(&any_kind, "(IS_NULL $0)"),
+            unary(&any_kind, "(IS_NOT_NULL $0)"),
+            binary(&n, &n, "(EQ $0 $1)"),
+            binary(&s, &s, "(EQ $0 $1)"),
+            binary(&b, &b, "(EQ $0 $1)"),
+            binary(&n, &n, "(NEQ $0 $1)"),
+            binary(&s, &s, "(NEQ $0 $1)"),
+            binary(&n, &n, "(LT $0 $1)"),
+            binary(&n, &n, "(LT_EQ $0 $1)"),
+            binary(&n, &n, "(GT $0 $1)"),
+            binary(&s, &s, "(GT_EQ $0 $1)"),
+            unary(&s, r#"(REGEXP_LIKE $0 "a")"#),
+            ternary(&b, &b, &b, "(IF $0 $1 $2)"),
+        ])
+        .boxed();
+        let number = prop::strategy::Union::new(vec![
+            leaves.number,
+            unary(&n, "(ADD $0)"),
+            binary(&n, &n, "(ADD $0 $1)"),
+            binary(&n, &n, "(SUBTRACT $0 $1)"),
+            binary(&n, &n, "(MULTIPLY $0 $1)"),
+            binary(&n, &n, "(DIVIDE $0 $1)"),
+            binary(&n, &n, "(POW $0 $1)"),
+            binary(&n, &n, "(INT_MOD $0 $1)"),
+            binary(&n, &n, "(MIN $0 $1)"),
+            binary(&n, &n, "(MAX $0 $1)"),
+            unary(&n, "(ABS $0)"),
+            unary(&n, "(CEIL $0)"),
+            unary(&n, "(FLOOR $0)"),
+            unary(&n, "(SQRT $0)"),
+            unary(&n, "(ROUND $0)"),
+            unary(&n, "(ROUND $0 1i64)"),
+            // SUBSTRING_COUNT is not generated: its native implementation builds a slice from a
+            // null pointer when the haystack is null, which aborts debug builds.
+            ternary(&b, &n, &n, "(IF $0 $1 $2)"),
+        ])
+        .boxed();
+        let string = prop::strategy::Union::new(vec![
+            leaves.string,
+            unary(&s, "(UPPER $0)"),
+            unary(&s, "(LOWER $0)"),
+            unary(&s, "(LEFT $0 1i64)"),
+            unary(&s, "(RIGHT $0 1i64)"),
+            unary(&s, "(SUBSTRING $0 0i64 1i64)"),
+            binary(&s, &s, r#"(CONCAT "," "false" $0 $1)"#),
+            binary(&s, &s, r#"(TEXT_JOIN "," "true" $0 $1)"#),
+            unary(&s, r#"(TRIM $0 "a" "both")"#),
+            unary(&s, r#"(SPLIT_AFTER $0 ",")"#),
+            unary(&s, r#"(SPLIT_BEFORE $0 "," 0i64)"#),
+            unary(&s, r#"(REGEXP_EXTRACT $0 "(a)b" 1u64)"#),
+            // IF is not generated for strings: with a null condition, it returns the selected
+            // branch instead of null, as the string pointer is not cleared.
+        ])
+        .boxed();
+        ExprStrategies {
+            boolean,
+            number,
+            string,
+        }
+    }
+
+    /// Compiles `expr_str`, then checks that `required(expr)` holds for every assignment where
+    /// `holds(result)` is true.
+    ///
+    /// Following tantivy's fast field binding, a variable is bound only if its type is accepted by
+    /// type inference. Unbound variables are null, and therefore absent.
+    fn check_necessary_condition(
+        expr_str: &str,
+        target_type: InferredTypeSet,
+        assignments: &[Assignment],
+        required: fn(&UntypedExpr) -> VariablePresenceCondition,
+        holds: fn(VarType, VariableValue) -> bool,
+    ) -> Result<(), TestCaseError> {
+        let expr = deserialize(expr_str).unwrap();
+        let Ok(inferred_types) = infer_types_with_target(&expr, target_type) else {
+            return Err(TestCaseError::reject("type inference failed"));
+        };
+        let mut variable_types: HashMap<&str, VarType> =
+            HashMap::with_capacity(inferred_types.len());
+        for (variable_name, accepted_types) in &inferred_types {
+            let (_, var_type) = VARIABLES
+                .iter()
+                .find(|(name, _)| name == variable_name)
+                .unwrap();
+            if accepted_types.contains(*var_type) {
+                variable_types.insert(*variable_name, *var_type);
+            }
+        }
+        // Some expressions trip debug assertions of the compiler, unrelated to presence. For
+        // instance, `(SQRT (CEIL n0))` asks CEIL for a f64, while it always returns an i64.
+        let compile_result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+            compile(&expr, &variable_types)
+        }));
+        let Ok(Ok(compiled_fn)) = compile_result else {
+            return Err(TestCaseError::reject("compilation failed"));
+        };
+        let condition = required(&expr);
+        let mut string_arena = StringArena::default();
+        for assignment in assignments {
+            let variable_ord = |variable_name: &str| {
+                VARIABLES
+                    .iter()
+                    .position(|(name, _)| *name == variable_name)
+                    .unwrap()
+            };
+            let args: Vec<VariableValue> = compiled_fn
+                .inputs()
+                .iter()
+                .map(
+                    |input| match assignment[variable_ord(&input.variable_name)] {
+                        Some(value_ord) => variable_value(input.r#type, value_ord),
+                        None => VariableValue::none(),
+                    },
+                )
+                .collect();
+            // SAFETY: Each slot follows the compiled input order, and uses the input type.
+            let result = unsafe { compiled_fn.call(&args, &mut string_arena) };
+            if !holds(compiled_fn.result_type(), result) {
+                continue;
+            }
+            let mut is_present = |variable_name: &str| {
+                variable_types.contains_key(variable_name)
+                    && assignment[variable_ord(variable_name)].is_some()
+            };
+            prop_assert!(
+                condition.eval(&mut is_present),
+                "{expr_str} holds for {assignment:?}, but {condition:?} does not"
+            );
+        }
+        Ok(())
+    }
+
+    fn is_true(result_type: VarType, result: VariableValue) -> bool {
+        // SAFETY: The union member is selected with the result type.
+        result_type == VarType::Bool && unsafe { result.as_bool() } == Some(true)
+    }
+
+    fn is_present(result_type: VarType, result: VariableValue) -> bool {
+        // SAFETY: The union member is selected with the result type.
+        unsafe {
+            match result_type {
+                VarType::Bool => result.as_bool().is_some(),
+                VarType::F64 => result.as_f64().is_some(),
+                VarType::U64 => result.as_u64().is_some(),
+                VarType::I64 => result.as_i64().is_some(),
+                VarType::Str => result.as_str().is_some(),
+                VarType::None => false,
+            }
+        }
+    }
+
+    proptest! {
+        // Compiler debug assertions reject a fraction of the generated expressions.
+        #![proptest_config(ProptestConfig {
+            max_global_rejects: 1 << 16,
+            ..ProptestConfig::with_cases(512)
+        })]
+
+        #[test]
+        fn proptest_required_presence_for_true_is_necessary(
+            expr in exprs(3).boolean,
+            assignments in assignments(),
+        ) {
+            check_necessary_condition(
+                &expr,
+                InferredTypeSet::BOOLEAN,
+                &assignments,
+                required_presence_for_true,
+                is_true,
+            )?;
+        }
+
+        #[test]
+        fn proptest_required_presence_is_necessary(
+            expr in prop_oneof![exprs(3).boolean, exprs(3).number, exprs(3).string],
+            assignments in assignments(),
+        ) {
+            check_necessary_condition(
+                &expr,
+                InferredTypeSet::ALL,
+                &assignments,
+                required_presence,
+                is_present,
+            )?;
+        }
+    }
+}
diff --git a/jitexpr/src/ast/serde.rs b/jitexpr/src/ast/serde.rs
index 705ad8d6c..959edff15 100644
--- a/jitexpr/src/ast/serde.rs
+++ b/jitexpr/src/ast/serde.rs
@@ -27,6 +27,7 @@ use std::fmt;
 use std::sync::Arc;
 
 use crate::ast::{Function, Literal, UntypedExpr};
+use crate::types::SafeF64;
 
 /// Serializes an untyped expression into its canonical Lisp-like form.
 pub fn serialize(expr: &UntypedExpr) -> String {
@@ -118,7 +119,7 @@ fn format_literal(literal: &Literal, formatter: &mut fmt::Formatter) -> fmt::Res
         Literal::Bool(value) => write!(formatter, "{value}"),
         Literal::U64(value) => write!(formatter, "{value}u64"),
         Literal::I64(value) => write!(formatter, "{value}i64"),
-        Literal::F64(value) => write!(formatter, "{value}f64"),
+        Literal::F64(value) => write!(formatter, "{}f64", value.get()),
         Literal::String(value) => format_quoted(value, '"', formatter),
     }
 }
@@ -273,7 +274,9 @@ fn parse_literal_atom(atom: &str) -> Option<Literal> {
     }
     if let Some(value_str) = atom.strip_suffix("f64") {
         let val = value_str.parse::<f64>().ok()?;
-        return Some(Literal::F64(val));
+        // Yields `None` for a non-finite float, which `parse_atom` reports as
+        // an error rather than letting it fall through to a variable name.
+        return SafeF64::new(val).map(Literal::F64);
     }
     None
 }
@@ -372,16 +375,19 @@ impl<'a> Parser<'a> {
         let atom_offset = self.offset;
         let atom = self.take_atom();
         if let Some(literal) = parse_literal_atom(atom) {
-            if let Literal::F64(value) = &literal
-                && !value.is_finite()
-            {
-                return Err(DeserializeError::new(
-                    atom_offset,
-                    format!("f64 literal `{atom}` must be finite"),
-                ));
-            }
             return Ok(UntypedExpr::Literal(literal));
         }
+        // An `f64`-suffixed atom that parses as a float but produced no literal
+        // is non-finite: `SafeF64` cannot hold it, and it must not be mistaken
+        // for a variable name.
+        if let Some(value_str) = atom.strip_suffix("f64")
+            && value_str.parse::<f64>().is_ok()
+        {
+            return Err(DeserializeError::new(
+                atom_offset,
+                format!("f64 literal `{atom}` must be finite"),
+            ));
+        }
         Ok(UntypedExpr::Variable(Arc::from(atom)))
     }
 
@@ -529,8 +535,14 @@ mod tests {
             (UntypedExpr::literal(false), "false"),
             (UntypedExpr::literal(u64::MAX), "18446744073709551615u64"),
             (UntypedExpr::literal(i64::MIN), "-9223372036854775808i64"),
-            (UntypedExpr::literal(1.5f64), "1.5f64"),
-            (UntypedExpr::literal(1.0f64), "1f64"),
+            (
+                UntypedExpr::literal(Literal::try_from(1.5f64).unwrap()),
+                "1.5f64",
+            ),
+            (
+                UntypedExpr::literal(Literal::try_from(1.0f64).unwrap()),
+                "1f64",
+            ),
         ];
 
         for (expr, expected) in cases {
@@ -579,19 +591,15 @@ mod tests {
 
     #[test]
     fn test_finite_float_edge_values_round_trip() {
-        for value in [
-            f64::MIN,
-            f64::MAX,
-            f64::MIN_POSITIVE,
-            f64::from_bits(1),
-            -0.0,
-        ] {
-            let serialized = serialize(&UntypedExpr::literal(value));
+        // `-0.0` is deliberately absent: `SafeF64` folds it to `0.0`, so it
+        // serializes as `0f64` and cannot round-trip.
+        for value in [f64::MIN, f64::MAX, f64::MIN_POSITIVE, f64::from_bits(1)] {
+            let serialized = serialize(&UntypedExpr::literal(Literal::try_from(value).unwrap()));
             let UntypedExpr::Literal(Literal::F64(parsed)) = deserialize(&serialized).unwrap()
             else {
                 panic!("expected an f64 literal");
             };
-            assert_eq!(parsed.to_bits(), value.to_bits());
+            assert_eq!(parsed.get().to_bits(), value.to_bits());
         }
     }
 
diff --git a/jitexpr/src/compile/cache.rs b/jitexpr/src/compile/cache.rs
new file mode 100644
index 000000000..54326b027
--- /dev/null
+++ b/jitexpr/src/compile/cache.rs
@@ -0,0 +1,306 @@
+use std::collections::HashMap;
+use std::num::NonZeroUsize;
+use std::sync::{Arc, Mutex, OnceLock};
+
+use lru::LruCache;
+
+use super::{CompileError, CompiledFn};
+use crate::ast::UntypedExpr;
+use crate::types::VarType;
+
+/// The outcome of compiling one key, shared by every caller of that key.
+type CompilationResult = Result<Arc<CompiledFn>, CompileError>;
+
+/// A per-key cell, written once by whichever caller compiles the expression.
+///
+/// Initializing it is what serializes concurrent compilations of the same key.
+type CompilationSlot = Arc<OnceLock<CompilationResult>>;
+
+/// A bounded, thread-safe cache of JIT-compiled expressions.
+/// The cache is cheap to clone: every clone shares one set of entries.
+/// The cache does not allocate on creation. Allocation happens on the first usage.
+#[derive(Clone)]
+pub struct ExprCompilationCache {
+    inner: Arc<Mutex<ExprCompilationCacheInner>>,
+}
+
+struct ExprCompilationCacheInner {
+    capacity: usize,
+    // We use Option here to lazily allocate on the first insertion.
+    entries: Option<LruCache<ExprCacheKey, CompilationSlot>>,
+}
+
+impl ExprCompilationCacheInner {
+    fn entries(&mut self) -> Option<&mut LruCache<ExprCacheKey, CompilationSlot>> {
+        if self.entries.is_none() {
+            let non_zero_capacity = NonZeroUsize::new(self.capacity)?;
+            self.entries = Some(LruCache::new(non_zero_capacity));
+        }
+        self.entries.as_mut()
+    }
+}
+
+/// Identifies a compilation: an expression plus the types it was compiled for.
+#[derive(PartialEq, Eq, Hash)]
+struct ExprCacheKey {
+    expr: String,
+    /// The variable types, sorted by variable name so the key does not depend
+    /// on the caller's `HashMap` iteration order.
+    var_types: Box<[(String, VarType)]>,
+}
+
+impl ExprCacheKey {
+    fn new(untyped_expr: &UntypedExpr, var_types: &HashMap<&str, VarType>) -> ExprCacheKey {
+        let mut sorted_var_types: Vec<(String, VarType)> = Vec::with_capacity(var_types.len());
+        for (variable_name, var_type) in var_types {
+            sorted_var_types.push((variable_name.to_string(), *var_type));
+        }
+        sorted_var_types.sort_unstable();
+        ExprCacheKey {
+            expr: untyped_expr.to_string(),
+            var_types: sorted_var_types.into_boxed_slice(),
+        }
+    }
+}
+
+impl ExprCompilationCache {
+    /// A capacity of 0 means disabled.
+    pub fn with_capacity(capacity: usize) -> ExprCompilationCache {
+        ExprCompilationCache {
+            inner: Arc::new(Mutex::new(ExprCompilationCacheInner {
+                capacity,
+                entries: None,
+            })),
+        }
+    }
+
+    /// Creates a cache that memoizes nothing and allocates nothing.
+    pub fn disabled() -> ExprCompilationCache {
+        ExprCompilationCache::with_capacity(0)
+    }
+
+    /// Returns false for a cache that memoizes nothing.
+    pub fn is_enabled(&self) -> bool {
+        self.inner.lock().unwrap().capacity > 0
+    }
+
+    /// Returns the number of compiled expressions currently retained.
+    pub fn len(&self) -> usize {
+        let mut inner_guard = self.inner.lock().unwrap();
+        let Some(entries) = inner_guard.entries() else {
+            return 0;
+        };
+        entries.len()
+    }
+
+    /// Returns true if the cache retains no compiled expression.
+    pub fn is_empty(&self) -> bool {
+        self.len() == 0
+    }
+
+    /// Returns the expression compiled for `var_types`, compiling it on a miss.
+    ///
+    /// Concurrent callers asking for the same expression and types block until
+    /// the first of them is done, so an expression is normally compiled once.
+    /// An expression evicted while a compilation is in flight may be compiled
+    /// again by a later caller.
+    ///
+    /// A compilation failure is cached like a success and handed to later
+    /// callers.
+    pub fn compile(
+        &self,
+        untyped_expr: &UntypedExpr,
+        var_types: &HashMap<&str, VarType>,
+    ) -> Result<Arc<CompiledFn>, CompileError> {
+        if let Some(slot) = self.slot_opt(untyped_expr, var_types) {
+            // Initializing the cell ensures we cannot have two threads compiling
+            // the same function at the same time.
+            slot.get_or_init(|| super::compile(untyped_expr, var_types))
+                .clone()
+        } else {
+            // no caching
+            super::compile(untyped_expr, var_types)
+        }
+    }
+
+    fn slot_opt(
+        &self,
+        untyped_expr: &UntypedExpr,
+        var_types: &HashMap<&str, VarType>,
+    ) -> Option<CompilationSlot> {
+        let key = ExprCacheKey::new(untyped_expr, var_types);
+        // That function does take the lock but only does trivial things that
+        // cannot panick before releasing it.
+        let mut inner_guard = self.inner.lock().unwrap();
+        let entries = inner_guard.entries()?;
+        let slot: CompilationSlot = entries.get_or_insert(key, CompilationSlot::default).clone();
+        Some(slot)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Barrier;
+
+    use super::*;
+    use crate::ast::Function;
+
+    #[test]
+    fn test_cache_hit_returns_the_same_compiled_fn() {
+        let cache = ExprCompilationCache::with_capacity(64);
+        let untyped_expr = UntypedExpr::variable("flag");
+        let variable_types = HashMap::from([("flag", VarType::Bool)]);
+        let first = cache.compile(&untyped_expr, &variable_types).unwrap();
+        let second = cache.compile(&untyped_expr, &variable_types).unwrap();
+        assert!(Arc::ptr_eq(&first, &second));
+        assert_eq!(cache.len(), 1);
+    }
+
+    #[test]
+    fn test_var_types_are_part_of_the_key() {
+        let cache = ExprCompilationCache::with_capacity(64);
+        let untyped_expr = UntypedExpr::variable("value");
+
+        let as_u64 = cache
+            .compile(&untyped_expr, &HashMap::from([("value", VarType::U64)]))
+            .unwrap();
+        let as_i64 = cache
+            .compile(&untyped_expr, &HashMap::from([("value", VarType::I64)]))
+            .unwrap();
+
+        assert!(!Arc::ptr_eq(&as_u64, &as_i64));
+        assert_eq!(as_u64.result_type(), VarType::U64);
+        assert_eq!(as_i64.result_type(), VarType::I64);
+        assert_eq!(cache.len(), 2);
+    }
+
+    #[test]
+    fn test_var_types_order_is_not_part_of_the_key() {
+        let untyped_expr = Function::Add
+            .call(vec![
+                UntypedExpr::variable("arg1"),
+                UntypedExpr::variable("arg2"),
+                UntypedExpr::variable("arg3"),
+                UntypedExpr::variable("arg4"),
+            ])
+            .unwrap();
+        let var_args = HashMap::from([
+            ("arg2", VarType::I64),
+            ("arg4", VarType::Str),
+            ("arg3", VarType::F64),
+            ("arg1", VarType::U64),
+        ]);
+        let key = ExprCacheKey::new(&untyped_expr, &var_args);
+        assert_eq!(key.var_types.len(), 4);
+        assert_eq!(key.var_types[0].0, "arg1");
+        assert_eq!(key.var_types[1].0, "arg2");
+        assert_eq!(key.var_types[2].0, "arg3");
+        assert_eq!(key.var_types[3].0, "arg4");
+    }
+
+    #[test]
+    fn test_equal_expressions_built_separately_share_an_entry() {
+        let cache = ExprCompilationCache::with_capacity(64);
+        let variable_types = HashMap::from([("value", VarType::U64)]);
+        let make_expr = || {
+            Function::Add
+                .call(vec![
+                    UntypedExpr::variable("value"),
+                    UntypedExpr::literal(1u64),
+                ])
+                .unwrap()
+        };
+
+        let first = cache.compile(&make_expr(), &variable_types).unwrap();
+        let second = cache.compile(&make_expr(), &variable_types).unwrap();
+
+        assert!(Arc::ptr_eq(&first, &second));
+        assert_eq!(cache.len(), 1);
+    }
+
+    #[test]
+    fn test_least_recently_used_entry_is_evicted() {
+        let cache = ExprCompilationCache::with_capacity(1);
+        let variable_types = HashMap::from([("flag", VarType::Bool)]);
+        let first_expr = UntypedExpr::variable("flag");
+        let second_expr = Function::Not
+            .call(vec![UntypedExpr::variable("flag")])
+            .unwrap();
+
+        let first = cache.compile(&first_expr, &variable_types).unwrap();
+        assert_eq!(cache.len(), 1);
+        cache.compile(&second_expr, &variable_types).unwrap();
+        assert_eq!(cache.len(), 1);
+        let first_again = cache.compile(&first_expr, &variable_types).unwrap();
+
+        assert!(!Arc::ptr_eq(&first, &first_again));
+        assert_eq!(cache.len(), 1);
+    }
+
+    #[test]
+    fn test_disabled_cache_memoizes_nothing() {
+        let cache = ExprCompilationCache::disabled();
+        let untyped_expr = UntypedExpr::variable("flag");
+        let variable_types = HashMap::from([("flag", VarType::Bool)]);
+
+        let first = cache.compile(&untyped_expr, &variable_types).unwrap();
+        let second = cache.compile(&untyped_expr, &variable_types).unwrap();
+
+        assert!(!cache.is_enabled());
+        assert!(!Arc::ptr_eq(&first, &second));
+        assert_eq!(cache.len(), 0);
+        assert!(cache.is_empty());
+    }
+
+    #[test]
+    fn test_capacity_0_means_disabled() {
+        assert!(ExprCompilationCache::with_capacity(1).is_enabled());
+        assert!(!ExprCompilationCache::with_capacity(0).is_enabled());
+        assert!(!ExprCompilationCache::disabled().is_enabled());
+    }
+
+    #[test]
+    fn test_compilation_failures_are_memoized() {
+        let cache = ExprCompilationCache::with_capacity(16);
+        // `(` is not a valid regular expression, which is rejected at compile time.
+        let untyped_expr = crate::ast::deserialize(r#"(REGEXP_EXTRACT "a" "*(" 1u64)"#).unwrap();
+        let variable_types = HashMap::new();
+        assert!(cache.is_empty());
+        assert!(cache.compile(&untyped_expr, &variable_types).is_err());
+        assert_eq!(cache.len(), 1);
+    }
+
+    #[test]
+    fn test_concurrent_callers_compile_once() {
+        const NUM_THREADS: usize = 8;
+
+        let cache = ExprCompilationCache::with_capacity(NUM_THREADS);
+        let barrier = Barrier::new(NUM_THREADS);
+        let untyped_expr = UntypedExpr::variable("flag");
+
+        let compiled_fns: Vec<Arc<CompiledFn>> = std::thread::scope(|scope| {
+            let handles: Vec<_> = (0..NUM_THREADS)
+                .map(|_| {
+                    let cache = cache.clone();
+                    let untyped_expr = &untyped_expr;
+                    let barrier = &barrier;
+                    scope.spawn(move || {
+                        let variable_types = HashMap::from([("flag", VarType::Bool)]);
+                        barrier.wait();
+                        cache.compile(untyped_expr, &variable_types).unwrap()
+                    })
+                })
+                .collect();
+            handles
+                .into_iter()
+                .map(|handle| handle.join().unwrap())
+                .collect()
+        });
+
+        // Identical pointers can only come from a single compilation.
+        for compiled_fn in &compiled_fns {
+            assert!(Arc::ptr_eq(&compiled_fns[0], compiled_fn));
+        }
+        assert_eq!(cache.len(), 1);
+    }
+}
diff --git a/jitexpr/src/compile/compile_fn_builder.rs b/jitexpr/src/compile/compile_fn_builder.rs
index 1df012ee5..cf6a870fa 100644
--- a/jitexpr/src/compile/compile_fn_builder.rs
+++ b/jitexpr/src/compile/compile_fn_builder.rs
@@ -15,7 +15,7 @@ use super::{
 };
 use crate::ast::{InferredTypeSet, Literal, UntypedExpr};
 use crate::functions::{declare_native_functions, register_jit_symbols};
-use crate::types::VarType;
+use crate::types::{SafeF64, VarType};
 
 pub(crate) struct CompileFnBuilder<'types, 'names> {
     variable_types: &'types HashMap<&'names str, VarType>,
@@ -109,8 +109,8 @@ impl<'types, 'names> CompileFnBuilder<'types, 'names> {
             match literal {
                 Literal::U64(value) => TypedLiteral::I64(*value as i64),
                 Literal::I64(value) => TypedLiteral::I64(*value),
-                Literal::F64(value) if f64_to_i64_lossless(*value).is_some() => {
-                    TypedLiteral::I64(f64_to_i64_lossless(*value).unwrap())
+                Literal::F64(value) if f64_to_i64_lossless(value.get()).is_some() => {
+                    TypedLiteral::I64(f64_to_i64_lossless(value.get()).unwrap())
                 }
                 _ => panic!("cannot coerce literal {literal:?} to i64"),
             }
@@ -118,15 +118,15 @@ impl<'types, 'names> CompileFnBuilder<'types, 'names> {
             match literal {
                 Literal::U64(value) => TypedLiteral::U64(*value),
                 Literal::I64(value) => TypedLiteral::U64(*value as u64),
-                Literal::F64(value) if f64_to_u64_lossless(*value).is_some() => {
-                    TypedLiteral::U64(f64_to_u64_lossless(*value).unwrap())
+                Literal::F64(value) if f64_to_u64_lossless(value.get()).is_some() => {
+                    TypedLiteral::U64(f64_to_u64_lossless(value.get()).unwrap())
                 }
                 _ => panic!("cannot coerce literal {literal:?} to u64"),
             }
         } else if intersection.contains(VarType::F64) {
             match literal {
-                Literal::U64(value) => TypedLiteral::F64(*value as f64),
-                Literal::I64(value) => TypedLiteral::F64(*value as f64),
+                Literal::U64(value) => TypedLiteral::F64(SafeF64::from_integer(*value)),
+                Literal::I64(value) => TypedLiteral::F64(SafeF64::from_integer(*value)),
                 Literal::F64(value) => TypedLiteral::F64(*value),
                 _ => panic!("cannot coerce literal {literal:?} to f64"),
             }
diff --git a/jitexpr/src/compile/error.rs b/jitexpr/src/compile/error.rs
index 2e6a48110..76f9339b9 100644
--- a/jitexpr/src/compile/error.rs
+++ b/jitexpr/src/compile/error.rs
@@ -1,12 +1,14 @@
+use std::sync::Arc;
+
 use crate::ast::{Function, InvalidFnCall, TypeError};
 use crate::types::VarType;
 
-#[derive(Debug, thiserror::Error)]
+#[derive(Debug, Clone, thiserror::Error)]
 pub enum CompileError {
     #[error("type inference failed: {0}")]
     TypeInference(#[from] TypeError),
     #[error("JIT compilation failed: {0}")]
-    Module(#[source] Box<cranelift_module::ModuleError>),
+    Module(#[source] Arc<cranelift_module::ModuleError>),
     #[error("cannot coerce an expression from {from_type:?} to {target:?}")]
     UnsupportedCoercion { from_type: VarType, target: VarType },
     #[error("cannot compile {function:?} with result type {return_type:?}")]
@@ -26,6 +28,6 @@ pub enum CompileError {
 
 impl From<cranelift_module::ModuleError> for CompileError {
     fn from(error: cranelift_module::ModuleError) -> Self {
-        CompileError::Module(Box::new(error))
+        CompileError::Module(Arc::new(error))
     }
 }
diff --git a/jitexpr/src/compile/mod.rs b/jitexpr/src/compile/mod.rs
index 7a2051e4f..65b83ec70 100644
--- a/jitexpr/src/compile/mod.rs
+++ b/jitexpr/src/compile/mod.rs
@@ -1,3 +1,4 @@
+mod cache;
 mod compile_fn_builder;
 mod compiled_fn;
 mod error;
@@ -8,6 +9,7 @@ mod typed_expr_serialize;
 use std::collections::HashMap;
 use std::sync::Arc;
 
+pub use cache::ExprCompilationCache;
 pub(crate) use compile_fn_builder::CompileFnBuilder;
 pub use compiled_fn::{CompiledFn, CompiledFnCtx};
 use cranelift::codegen::ir::{
@@ -170,7 +172,7 @@ fn lower_literal(
             builder
                 .ins()
                 .f64const(cranelift::codegen::ir::immediates::Ieee64::with_bits(
-                    value.to_bits(),
+                    value.get().to_bits(),
                 ))
         }
         TypedLiteral::String(value) => {
diff --git a/jitexpr/src/compile/typed_expr.rs b/jitexpr/src/compile/typed_expr.rs
index e0507040b..3b81d0ffb 100644
--- a/jitexpr/src/compile/typed_expr.rs
+++ b/jitexpr/src/compile/typed_expr.rs
@@ -3,7 +3,7 @@ use std::sync::Arc;
 #[cfg(test)]
 use crate::ast::Literal;
 use crate::functions::FnCallEnum;
-use crate::types::VarType;
+use crate::types::{SafeF64, VarType};
 
 #[derive(Clone, PartialEq)]
 pub struct TypedVariable {
@@ -34,29 +34,27 @@ impl TypedExpr {
                     TypedExprAst::Literal(TypedLiteral::I64(value as i64))
                 }
                 (TypedExprAst::Literal(TypedLiteral::U64(value)), VarType::F64) => {
-                    TypedExprAst::Literal(TypedLiteral::F64(value as f64))
+                    TypedExprAst::Literal(TypedLiteral::F64(SafeF64::from_integer(value)))
                 }
                 (TypedExprAst::Literal(TypedLiteral::I64(value)), VarType::U64) if value >= 0 => {
                     TypedExprAst::Literal(TypedLiteral::U64(value as u64))
                 }
                 (TypedExprAst::Literal(TypedLiteral::I64(value)), VarType::F64) => {
-                    TypedExprAst::Literal(TypedLiteral::F64(value as f64))
+                    TypedExprAst::Literal(TypedLiteral::F64(SafeF64::from_integer(value)))
                 }
                 (TypedExprAst::Literal(TypedLiteral::F64(value)), VarType::U64)
-                    if value.is_finite()
-                        && value.fract() == 0.0
-                        && value >= 0.0
-                        && value < u64::MAX as f64 =>
+                    if value.get().fract() == 0.0
+                        && value.get() >= 0.0
+                        && value.get() < u64::MAX as f64 =>
                 {
-                    TypedExprAst::Literal(TypedLiteral::U64(value as u64))
+                    TypedExprAst::Literal(TypedLiteral::U64(value.get() as u64))
                 }
                 (TypedExprAst::Literal(TypedLiteral::F64(value)), VarType::I64)
-                    if value.is_finite()
-                        && value.fract() == 0.0
-                        && value >= i64::MIN as f64
-                        && value < -(i64::MIN as f64) =>
+                    if value.get().fract() == 0.0
+                        && value.get() >= i64::MIN as f64
+                        && value.get() < -(i64::MIN as f64) =>
                 {
-                    TypedExprAst::Literal(TypedLiteral::I64(value as i64))
+                    TypedExprAst::Literal(TypedLiteral::I64(value.get() as i64))
                 }
                 (ast, target_type) => TypedExprAst::Coerce {
                     target_type,
@@ -99,7 +97,7 @@ pub(crate) enum TypedLiteral {
     Bool(bool),
     U64(u64),
     I64(i64),
-    F64(f64),
+    F64(SafeF64),
     String(Arc<str>),
 }
 
diff --git a/jitexpr/src/compile/typed_expr_serialize.rs b/jitexpr/src/compile/typed_expr_serialize.rs
index dbc6b7571..65a431c77 100644
--- a/jitexpr/src/compile/typed_expr_serialize.rs
+++ b/jitexpr/src/compile/typed_expr_serialize.rs
@@ -46,7 +46,7 @@ fn format_literal(literal: &TypedLiteral, formatter: &mut fmt::Formatter) -> fmt
         TypedLiteral::Bool(value) => write!(formatter, "{value}"),
         TypedLiteral::U64(value) => write!(formatter, "{value}u64"),
         TypedLiteral::I64(value) => write!(formatter, "{value}i64"),
-        TypedLiteral::F64(value) => write!(formatter, "{value}f64"),
+        TypedLiteral::F64(value) => write!(formatter, "{}f64", value.get()),
         TypedLiteral::String(value) => format_string_literal(value, formatter),
     }
 }
diff --git a/jitexpr/src/functions/add.rs b/jitexpr/src/functions/add.rs
index 6182ae397..740af0265 100644
--- a/jitexpr/src/functions/add.rs
+++ b/jitexpr/src/functions/add.rs
@@ -181,7 +181,10 @@ mod tests {
     fn test_infer_types_rejects_string_argument() {
         let expr = UntypedExpr::new_fn_call(
             Function::Add,
-            vec![UntypedExpr::literal(1.0), UntypedExpr::literal("hello")],
+            vec![
+                UntypedExpr::literal(Literal::try_from(1.0).unwrap()),
+                UntypedExpr::literal("hello"),
+            ],
         )
         .unwrap();
         let error = infer_types(&expr).unwrap_err();
@@ -387,7 +390,7 @@ mod tests {
             vec![
                 UntypedExpr::variable("myfield"),
                 UntypedExpr::literal(-2i64),
-                UntypedExpr::literal(0.5f64),
+                UntypedExpr::literal(Literal::try_from(0.5f64).unwrap()),
             ],
         )
         .unwrap();
@@ -428,7 +431,10 @@ mod tests {
     fn test_compile_u64_to_float_coercion_is_unsigned() {
         let expression = UntypedExpr::new_fn_call(
             Function::Add,
-            vec![UntypedExpr::variable("x"), UntypedExpr::literal(0.5f64)],
+            vec![
+                UntypedExpr::variable("x"),
+                UntypedExpr::literal(Literal::try_from(0.5f64).unwrap()),
+            ],
         )
         .unwrap();
         let variable_types = HashMap::from([("x", VarType::U64)]);
@@ -467,7 +473,10 @@ mod tests {
     fn test_compile_can_coerce_variable_when_necessary() {
         let expression = UntypedExpr::new_fn_call(
             Function::Add,
-            vec![UntypedExpr::variable("x"), UntypedExpr::literal(1.2f64)],
+            vec![
+                UntypedExpr::variable("x"),
+                UntypedExpr::literal(Literal::try_from(1.2f64).unwrap()),
+            ],
         )
         .unwrap();
         let variable_types = HashMap::from([("x", VarType::U64)]);
@@ -497,7 +506,10 @@ mod tests {
 
     #[test]
     fn test_no_variable_works() {
-        let args = vec![UntypedExpr::literal(1.2f64), UntypedExpr::literal(1u64)];
+        let args = vec![
+            UntypedExpr::literal(Literal::try_from(1.2f64).unwrap()),
+            UntypedExpr::literal(1u64),
+        ];
         let variable_types = HashMap::new();
         let typed_expr = crate::typed_expr_from_str("(ADD 1.2f64 1u64)", &variable_types);
         assert_eq!(typed_expr.return_type, VarType::F64);
diff --git a/jitexpr/src/functions/is_null.rs b/jitexpr/src/functions/is_null.rs
index b7bf84d68..3ef9645fa 100644
--- a/jitexpr/src/functions/is_null.rs
+++ b/jitexpr/src/functions/is_null.rs
@@ -98,7 +98,7 @@ impl From<IsNullFnCall> for FnCallEnum {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::ast::{InvalidFnCall, deserialize, infer_types};
+    use crate::ast::{InvalidFnCall, Literal, deserialize, infer_types};
     use crate::compile::compile;
     use crate::functions::ArgumentCount;
     use crate::types::VariableValue;
@@ -164,12 +164,25 @@ mod tests {
 
     #[test]
     fn test_non_finite_values_are_present() {
-        // The textual parser rejects non-finite literals, but programmatically
-        // constructed expressions can still contain them. They are not null.
-        for value in [f64::NAN, f64::INFINITY, f64::NEG_INFINITY] {
-            let expression =
-                UntypedExpr::new_fn_call(Function::IsNull, vec![UntypedExpr::literal(value)])
-                    .unwrap();
+        // `SafeF64` bars non-finite *literals*, but arithmetic still reaches
+        // non-finite *values* at runtime. Those are present, hence not null.
+        let float_literal = |value: f64| UntypedExpr::literal(Literal::try_from(value).unwrap());
+        let multiply = |left: UntypedExpr, right: UntypedExpr| {
+            UntypedExpr::new_fn_call(Function::Multiply, vec![left, right]).unwrap()
+        };
+
+        // `f64::MAX * f64::MAX` overflows to an infinity, and subtracting two
+        // like-signed infinities yields NaN.
+        let positive_infinity = multiply(float_literal(f64::MAX), float_literal(f64::MAX));
+        let negative_infinity = multiply(float_literal(f64::MIN), float_literal(f64::MAX));
+        let not_a_number = UntypedExpr::new_fn_call(
+            Function::Subtract,
+            vec![positive_infinity.clone(), positive_infinity.clone()],
+        )
+        .unwrap();
+
+        for non_finite in [positive_infinity, negative_infinity, not_a_number] {
+            let expression = UntypedExpr::new_fn_call(Function::IsNull, vec![non_finite]).unwrap();
             let mut compiled = compile(&expression, &HashMap::new()).unwrap().context();
             // SAFETY: The expression has no runtime inputs and returns a boolean.
             assert_eq!(unsafe { compiled.call(&[]).as_bool() }, Some(false));
diff --git a/jitexpr/src/functions/left.rs b/jitexpr/src/functions/left.rs
index f194314b6..d0ab71078 100644
--- a/jitexpr/src/functions/left.rs
+++ b/jitexpr/src/functions/left.rs
@@ -38,7 +38,7 @@ fn constant_length(expression: &UntypedExpr) -> Result<Option<usize>, super::Inv
     Ok(match literal {
         Literal::I64(value) => usize::try_from(*value).ok(),
         Literal::U64(value) => usize::try_from(*value).ok(),
-        Literal::F64(value) => usize::try_from(*value as i64).ok(),
+        Literal::F64(value) => usize::try_from(value.get() as i64).ok(),
         Literal::None => None,
         Literal::Bool(_) | Literal::String(_) => None,
     })
diff --git a/jitexpr/src/functions/mod.rs b/jitexpr/src/functions/mod.rs
index 734d46a99..1c49825b0 100644
--- a/jitexpr/src/functions/mod.rs
+++ b/jitexpr/src/functions/mod.rs
@@ -599,7 +599,7 @@ impl FnCallEnum {
 }
 
 /// Error representing an invalid function call.
-#[derive(Debug, Eq, PartialEq, thiserror::Error)]
+#[derive(Debug, Eq, PartialEq, thiserror::Error, Clone)]
 pub enum InvalidFnCall {
     #[error("invalid number of arguments: expected {expected}, got {provided}")]
     InvalidNumberOfArguments {
diff --git a/jitexpr/src/functions/right.rs b/jitexpr/src/functions/right.rs
index 1c4efb2fd..e2c638f9d 100644
--- a/jitexpr/src/functions/right.rs
+++ b/jitexpr/src/functions/right.rs
@@ -39,7 +39,7 @@ fn constant_length(expression: &UntypedExpr) -> Result<Option<usize>, super::Inv
     Ok(match literal {
         Literal::I64(value) => usize::try_from(*value).ok(),
         Literal::U64(value) => usize::try_from(*value).ok(),
-        Literal::F64(value) => usize::try_from(*value as i64).ok(),
+        Literal::F64(value) => usize::try_from(value.get() as i64).ok(),
         Literal::None => None,
         Literal::Bool(_) | Literal::String(_) => None,
     })
diff --git a/jitexpr/src/functions/round.rs b/jitexpr/src/functions/round.rs
index d526a2dee..28e0c1071 100644
--- a/jitexpr/src/functions/round.rs
+++ b/jitexpr/src/functions/round.rs
@@ -57,12 +57,11 @@ fn constant_precision(
         Literal::I64(value) => Some(*value),
         Literal::U64(value) => i64::try_from(*value).ok(),
         Literal::F64(value)
-            if value.is_finite()
-                && value.fract() == 0.0
-                && *value >= i64::MIN as f64
-                && *value < -(i64::MIN as f64) =>
+            if value.get().fract() == 0.0
+                && value.get() >= i64::MIN as f64
+                && value.get() < -(i64::MIN as f64) =>
         {
-            Some(*value as i64)
+            Some(value.get() as i64)
         }
         Literal::None => None,
         Literal::F64(_) | Literal::Bool(_) | Literal::String(_) => None,
diff --git a/jitexpr/src/functions/split_after.rs b/jitexpr/src/functions/split_after.rs
index 31705ecc9..5f0b3c0a9 100644
--- a/jitexpr/src/functions/split_after.rs
+++ b/jitexpr/src/functions/split_after.rs
@@ -52,7 +52,7 @@ fn constant_occurrence(
     Ok(match literal {
         Literal::I64(value) => usize::try_from(*value).ok(),
         Literal::U64(value) => usize::try_from(*value).ok(),
-        Literal::F64(value) => usize::try_from(*value as i64).ok(),
+        Literal::F64(value) => usize::try_from(value.get() as i64).ok(),
         Literal::None => None,
         Literal::Bool(_) | Literal::String(_) => None,
     })
diff --git a/jitexpr/src/functions/split_before.rs b/jitexpr/src/functions/split_before.rs
index 67415f6fd..4e75aadbc 100644
--- a/jitexpr/src/functions/split_before.rs
+++ b/jitexpr/src/functions/split_before.rs
@@ -51,7 +51,7 @@ fn constant_occurrence(
     Ok(match literal {
         Literal::I64(value) => usize::try_from(*value).ok(),
         Literal::U64(value) => usize::try_from(*value).ok(),
-        Literal::F64(value) => usize::try_from(*value as i64).ok(),
+        Literal::F64(value) => usize::try_from(value.get() as i64).ok(),
         Literal::None => None,
         Literal::Bool(_) | Literal::String(_) => None,
     })
diff --git a/jitexpr/src/functions/sqrt.rs b/jitexpr/src/functions/sqrt.rs
index b914f48ac..cadaf986b 100644
--- a/jitexpr/src/functions/sqrt.rs
+++ b/jitexpr/src/functions/sqrt.rs
@@ -145,8 +145,11 @@ mod tests {
         assert_eq!(eval("(SQRT none)"), None);
         assert_eq!(eval("(SQRT -1i64)"), None);
 
-        let negative_zero = eval("(SQRT -0f64)").unwrap();
-        assert_eq!(negative_zero.to_bits(), (-0.0f64).to_bits());
+        // `SafeF64` folds `-0.0` on construction, so the literal carries a
+        // positive zero and IEEE's `sqrt(-0.0) == -0.0` is unreachable here.
+        // The sign still exists for runtime values, as `abs.rs` exercises.
+        let zero = eval("(SQRT -0f64)").unwrap();
+        assert_eq!(zero.to_bits(), 0.0f64.to_bits());
     }
 
     #[test]
diff --git a/jitexpr/src/functions/substring.rs b/jitexpr/src/functions/substring.rs
index 5d2e2e92d..e15a6d18e 100644
--- a/jitexpr/src/functions/substring.rs
+++ b/jitexpr/src/functions/substring.rs
@@ -53,7 +53,7 @@ fn constant_usize(
             .ok()
             .and_then(|value| usize::try_from(value).ok()),
         Literal::F64(value) if literal.types().contains(VarType::I64) => {
-            usize::try_from(*value as i64).ok()
+            usize::try_from(value.get() as i64).ok()
         }
         Literal::None => None,
         Literal::Bool(_) | Literal::F64(_) | Literal::String(_) => None,
diff --git a/jitexpr/src/lib.rs b/jitexpr/src/lib.rs
index b6aa1517c..9d9d88dbf 100644
--- a/jitexpr/src/lib.rs
+++ b/jitexpr/src/lib.rs
@@ -1,3 +1,5 @@
+//! EXPERIMENTAL. The API is likely to change in the near future.
+
 pub mod ast;
 pub mod compile;
 pub mod types;
diff --git a/jitexpr/src/types.rs b/jitexpr/src/types.rs
index 18a8cb957..e767a0ce5 100644
--- a/jitexpr/src/types.rs
+++ b/jitexpr/src/types.rs
@@ -1,5 +1,9 @@
 //! Source types and nullable runtime value representations.
 
+use std::cmp::Ordering;
+use std::fmt;
+use std::hash::{Hash, Hasher};
+
 /// A value type supported by compiled expressions.
 #[derive(Clone, Copy, Debug, Eq, Hash, PartialEq, Ord, PartialOrd)]
 pub enum VarType {
@@ -11,6 +15,82 @@ pub enum VarType {
     None,
 }
 
+/// Wraps a f64 that is not inf, nor Nan, nor neg 0.
+#[derive(Copy, Clone)]
+pub struct SafeF64(f64);
+
+impl SafeF64 {
+    /// Returns `None` for NaN and for either infinity.
+    ///
+    /// A negative zero is accepted, and folded to `0.0`.
+    pub fn new(val: f64) -> Option<Self> {
+        if val.is_nan() || val.is_infinite() {
+            return None;
+        }
+        if val == 0.0 {
+            // True of both zeros; `0.0` is the representative.
+            Some(SafeF64(0.0))
+        } else {
+            Some(SafeF64(val))
+        }
+    }
+
+    /// Converts an integer to the nearest `f64`.
+    ///
+    /// Total, hence infallible: every `i64` and `u64` magnitude is far inside
+    /// `f64`'s finite range, so no non-finite value can come out. The
+    /// conversion may still round, exactly as `as f64` would.
+    pub fn from_integer(value: impl Into<i128>) -> SafeF64 {
+        SafeF64(value.into() as f64)
+    }
+
+    /// Returns the wrapped value, which is guaranteed finite.
+    pub fn get(self) -> f64 {
+        self.0
+    }
+}
+
+/// Forwards to the wrapped `f64`, so a `SafeF64` is indistinguishable from the
+/// number it holds.
+impl fmt::Debug for SafeF64 {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        fmt::Debug::fmt(&self.0, formatter)
+    }
+}
+
+impl PartialEq for SafeF64 {
+    fn eq(&self, other: &SafeF64) -> bool {
+        self.0 == other.0
+    }
+}
+
+// Sound because NaN is excluded, so equality is reflexive.
+impl Eq for SafeF64 {}
+
+impl Ord for SafeF64 {
+    #[inline(always)]
+    fn cmp(&self, other: &SafeF64) -> Ordering {
+        self.0
+            .partial_cmp(&other.0)
+            .expect("SafeF64 excludes NaN, so values are totally ordered")
+    }
+}
+
+impl PartialOrd for SafeF64 {
+    #[inline(always)]
+    fn partial_cmp(&self, other: &SafeF64) -> Option<Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+/// Hashes the bit pattern. Because we remove NaN Inf and -0.0,
+/// this is consistent with equality. (x == y => hash(x) == hash(y)).
+impl Hash for SafeF64 {
+    fn hash<H: Hasher>(&self, hasher: &mut H) {
+        self.0.to_bits().hash(hasher);
+    }
+}
+
 /// The payload of a primitive runtime value.
 ///
 /// This union is deliberately untagged. The corresponding
@@ -291,7 +371,121 @@ impl<'a> From<VariablePrimitiveOpt> for VariableValue<'a> {
 
 #[cfg(test)]
 mod tests {
-    use crate::types::{VariablePrimitive, VariablePrimitiveOpt, VariableValue};
+    use std::cmp::Ordering;
+    use std::collections::hash_map::DefaultHasher;
+    use std::hash::{Hash, Hasher};
+
+    use crate::types::{SafeF64, VariablePrimitive, VariablePrimitiveOpt, VariableValue};
+
+    fn safe(val: f64) -> SafeF64 {
+        SafeF64::new(val).unwrap()
+    }
+
+    fn hash_of(val: SafeF64) -> u64 {
+        let mut hasher = DefaultHasher::new();
+        val.hash(&mut hasher);
+        hasher.finish()
+    }
+
+    /// Values spanning both zeros -- which fold together -- and both extremes.
+    fn sample_values() -> Vec<SafeF64> {
+        [
+            -0.0f64,
+            0.0,
+            1.0,
+            -1.0,
+            1.5,
+            -1.5,
+            f64::MIN,
+            f64::MAX,
+            f64::MIN_POSITIVE,
+        ]
+        .into_iter()
+        .map(safe)
+        .collect()
+    }
+
+    #[test]
+    fn test_safe_f64_rejects_non_finite() {
+        assert!(SafeF64::new(f64::NAN).is_none());
+        assert!(SafeF64::new(-f64::NAN).is_none());
+        assert!(SafeF64::new(f64::INFINITY).is_none());
+        assert!(SafeF64::new(f64::NEG_INFINITY).is_none());
+        assert_eq!(SafeF64::new(1.5).unwrap().get(), 1.5);
+        assert_eq!(SafeF64::new(f64::MAX).unwrap().get(), f64::MAX);
+    }
+
+    #[test]
+    fn test_safe_f64_eq_and_ord_laws() {
+        let values = sample_values();
+        for left in &values {
+            // Reflexivity is what NaN would have broken, making `Eq` unsound.
+            assert_eq!(left, left);
+            assert_eq!(left.cmp(left), Ordering::Equal);
+            for right in &values {
+                assert_eq!(left == right, right == left, "symmetry");
+                assert_eq!(left.cmp(right), right.cmp(left).reverse(), "antisymmetry");
+                // `Ord` must agree with `Eq`, and `Hash` with both.
+                assert_eq!(
+                    left == right,
+                    left.cmp(right) == Ordering::Equal,
+                    "cmp agrees with eq"
+                );
+                if left == right {
+                    assert_eq!(hash_of(*left), hash_of(*right), "equal values hash equally");
+                }
+                for third in &values {
+                    if left == right && right == third {
+                        assert_eq!(left, third, "transitivity of eq");
+                    }
+                    if left <= right && right <= third {
+                        assert!(left <= third, "transitivity of ord");
+                    }
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn test_safe_f64_folds_negative_zero() {
+        // The fold happens on the way in, so there is a single zero to compare.
+        assert_eq!(safe(-0.0).get().to_bits(), 0.0f64.to_bits());
+        assert!(!safe(-0.0).get().is_sign_negative());
+        assert_eq!(
+            SafeF64::from_integer(0i64).get().to_bits(),
+            0.0f64.to_bits()
+        );
+
+        assert_eq!(safe(-0.0), safe(0.0));
+        assert_eq!(safe(-0.0).cmp(&safe(0.0)), Ordering::Equal);
+        assert_eq!(hash_of(safe(-0.0)), hash_of(safe(0.0)));
+    }
+
+    #[test]
+    fn test_safe_f64_sorts_in_numeric_order() {
+        let mut values = sample_values();
+        values.sort();
+        let sorted: Vec<f64> = values.iter().map(|value| value.get()).collect();
+        assert_eq!(
+            sorted,
+            vec![
+                f64::MIN,
+                -1.5,
+                -1.0,
+                -0.0,
+                0.0,
+                f64::MIN_POSITIVE,
+                1.0,
+                1.5,
+                f64::MAX
+            ]
+        );
+    }
+
+    #[test]
+    fn test_safe_f64_debug_is_transparent() {
+        assert_eq!(format!("{:?}", safe(1.5)), format!("{:?}", 1.5f64));
+    }
 
     #[test]
     fn test_runtime_value_layouts() {
diff --git a/query-grammar/src/query_grammar.rs b/query-grammar/src/query_grammar.rs
index aaa7800e8..3d0d58b64 100644
--- a/query-grammar/src/query_grammar.rs
+++ b/query-grammar/src/query_grammar.rs
@@ -325,11 +325,10 @@ fn exists(inp: &str) -> IResult<&str, UserInputLeaf> {
             multispace0,
             char('*'),
             peek(alt((
+                value("", multispace1),
                 value(
                     "",
-                    satisfy(|c: char| {
-                        c.is_whitespace() || (ESCAPE_IN_WORD.contains(&c) && c != '\\')
-                    }),
+                    satisfy(|c: char| ESCAPE_IN_WORD.contains(&c) && c != '\\'),
                 ),
                 eof,
             ))),
@@ -345,11 +344,10 @@ fn exists_precond(inp: &str) -> IResult<&str, (), ()> {
             multispace0,
             char('*'),
             peek(alt((
+                value("", multispace1),
                 value(
                     "",
-                    satisfy(|c: char| {
-                        c.is_whitespace() || (ESCAPE_IN_WORD.contains(&c) && c != '\\')
-                    }),
+                    satisfy(|c: char| ESCAPE_IN_WORD.contains(&c) && c != '\\'),
                 ),
                 eof,
             ))), // we need to check this isn't a wildcard query
@@ -687,12 +685,20 @@ fn set_infallible(mut inp: &str) -> JResult<&str, UserInputLeaf> {
             return Ok((inp, (res, errs)));
         }
         errs.append(&mut space_error);
-        // TODO
-        // here we do the assumption term_or_phrase_infallible always consume something if the
-        // first byte is not `)` or ' '. If it did not, we would end up looping.
 
         let (rest, (delim_term, mut err)) = simple_term_infallible("]")(inp)?;
         errs.append(&mut err);
+        if rest.len() == inp.len() {
+            errs.push(LenientErrorInternal {
+                pos: inp.len(),
+                message: "missing ]".to_string(),
+            });
+            let res = UserInputLeaf::Set {
+                field: None,
+                elements,
+            };
+            return Ok((inp, (res, errs)));
+        }
         if let Some((_, term)) = delim_term {
             elements.push(term);
         }
@@ -1125,11 +1131,14 @@ pub fn parse_to_ast(inp: &str) -> IResult<&str, UserInputAst> {
 }
 
 pub fn parse_to_ast_lenient(query_str: &str) -> (UserInputAst, Vec<LenientError>) {
-    if query_str.trim().is_empty() {
+    if query_str
+        .chars()
+        .all(|c| matches!(c, ' ' | '\t' | '\r' | '\n'))
+    {
         return (UserInputAst::Clause(Vec::new()), Vec::new());
     }
     let (left, (res, mut errors)) = ast_infallible(query_str).unwrap();
-    if !left.trim().is_empty() {
+    if !left.is_empty() {
         errors.push(LenientErrorInternal {
             pos: left.len(),
             message: "unparsed end of query".to_string(),
diff --git a/query-grammar/src/user_input_ast.rs b/query-grammar/src/user_input_ast.rs
index b607e56cd..0451b244d 100644
--- a/query-grammar/src/user_input_ast.rs
+++ b/query-grammar/src/user_input_ast.rs
@@ -47,8 +47,9 @@ impl UserInputLeaf {
                 upper,
             },
             UserInputLeaf::Set { field: _, elements } => UserInputLeaf::Set { field, elements },
-            UserInputLeaf::Exists { field: _ } => UserInputLeaf::Exists {
-                field: field.expect("Exist query without a field isn't allowed"),
+            UserInputLeaf::Exists { field: _ } => match field {
+                Some(field) => UserInputLeaf::Exists { field },
+                None => UserInputLeaf::All,
             },
             UserInputLeaf::Regex { field: _, pattern } => UserInputLeaf::Regex { field, pattern },
         }
diff --git a/src/aggregation/accessor_helpers.rs b/src/aggregation/accessor_helpers.rs
index fa51041e4..9d050d1b9 100644
--- a/src/aggregation/accessor_helpers.rs
+++ b/src/aggregation/accessor_helpers.rs
@@ -1,10 +1,12 @@
 //! This will enhance the request tree with access to the fastfield and metadata.
 
 use std::io;
+use std::sync::Arc;
 
-use columnar::{Column, ColumnType};
+use columnar::{Column, ColumnType, DynamicColumn, DynamicColumnHandle};
 
-use crate::aggregation::{f64_to_fastfield_u64, Key};
+use crate::aggregation::value_source::ValueSource;
+use crate::aggregation::{f64_to_fastfield_u64, Key, ValueSourceRegistry};
 use crate::index::SegmentReader;
 
 /// Get the missing value as internal u64 representation
@@ -55,14 +57,38 @@ pub(crate) fn get_numeric_or_date_column_types() -> &'static [ColumnType] {
     ]
 }
 
-/// Get fast field reader or empty as default.
-pub(crate) fn get_ff_reader(
+fn resolve_registered_source(
     reader: &SegmentReader,
+    value_sources: &ValueSourceRegistry,
+    field_name: &str,
+    allowed_column_types_opt: Option<&[ColumnType]>,
+) -> crate::Result<Option<Arc<dyn ValueSource>>> {
+    let Some(provider) = value_sources.get(field_name) else {
+        return Ok(None);
+    };
+    let source = provider.for_segment(reader)?;
+    let column_type = source.column_type();
+    if let Some(allowed_column_types) = allowed_column_types_opt {
+        if !allowed_column_types.contains(&column_type) {
+            return Ok(None);
+        }
+    }
+    Ok(Some(source))
+}
+
+pub(crate) fn get_value_source(
+    reader: &SegmentReader,
+    value_sources: &ValueSourceRegistry,
     field_name: &str,
     allowed_column_types: Option<&[ColumnType]>,
-) -> crate::Result<(columnar::Column<u64>, ColumnType)> {
+) -> crate::Result<Arc<dyn ValueSource>> {
+    if let Some(registered) =
+        resolve_registered_source(reader, value_sources, field_name, allowed_column_types)?
+    {
+        return Ok(registered);
+    }
     let ff_fields = reader.fast_fields();
-    let ff_field_with_type = ff_fields
+    let (column, column_type) = ff_fields
         .u64_lenient_for_type(allowed_column_types, field_name)?
         .unwrap_or_else(|| {
             (
@@ -70,36 +96,52 @@ pub(crate) fn get_ff_reader(
                 ColumnType::U64,
             )
         });
-    Ok(ff_field_with_type)
+    // The empty-column shim stays physical on purpose: several fast paths check
+    // `as_column()` and would otherwise degrade for a merely absent field.
+    Ok(Arc::new((column, column_type)))
 }
 
 pub(crate) fn get_dynamic_columns(
     reader: &SegmentReader,
     field_name: &str,
 ) -> crate::Result<Vec<columnar::DynamicColumn>> {
-    let ff_fields = reader.fast_fields().dynamic_column_handles(field_name)?;
-    let cols = ff_fields
+    let dyn_col_handles: Vec<DynamicColumnHandle> =
+        reader.fast_fields().dynamic_column_handles(field_name)?;
+    let dyn_cols: Vec<DynamicColumn> = dyn_col_handles
         .iter()
-        .map(|h| h.open())
+        .map(DynamicColumnHandle::open)
         .collect::<io::Result<_>>()?;
-    assert!(!ff_fields.is_empty(), "field {field_name} not found");
-    Ok(cols)
+    assert!(!dyn_cols.is_empty(), "field {field_name} not found");
+    Ok(dyn_cols)
 }
 
-/// Get all fast field reader or empty as default.
+/// Get all block_value_sources or empty as default.
 ///
 /// Is guaranteed to return at least one column.
-pub(crate) fn get_all_ff_reader_or_empty(
+pub(crate) fn get_all_value_sources(
     reader: &SegmentReader,
+    value_sources: &ValueSourceRegistry,
     field_name: &str,
     allowed_column_types: Option<&[ColumnType]>,
     fallback_type: ColumnType,
-) -> crate::Result<Vec<(columnar::Column<u64>, ColumnType)>> {
+) -> crate::Result<Vec<Arc<dyn ValueSource>>> {
+    // A registered source shadows the physical type fan-out entirely.
+    if let Some(registered) =
+        resolve_registered_source(reader, value_sources, field_name, allowed_column_types)?
+    {
+        return Ok(vec![registered]);
+    }
     let ff_fields = reader.fast_fields();
-    let mut ff_field_with_type =
+    let mut ff_field_with_type: Vec<(Column, ColumnType)> =
         ff_fields.u64_lenient_for_type_all(allowed_column_types, field_name)?;
     if ff_field_with_type.is_empty() {
         ff_field_with_type.push((Column::build_empty_column(reader.num_docs()), fallback_type));
     }
-    Ok(ff_field_with_type)
+    Ok(ff_field_with_type
+        .into_iter()
+        .map(|(column, column_type)| {
+            let source: Arc<dyn ValueSource> = Arc::new((column, column_type));
+            source
+        })
+        .collect())
 }
diff --git a/src/aggregation/agg_data.rs b/src/aggregation/agg_data.rs
index 475cef2ea..0213e002e 100644
--- a/src/aggregation/agg_data.rs
+++ b/src/aggregation/agg_data.rs
@@ -7,8 +7,8 @@ use serde::Serialize;
 use tantivy_fst::Regex;
 
 use crate::aggregation::accessor_helpers::{
-    get_all_ff_reader_or_empty, get_dynamic_columns, get_ff_reader, get_missing_val_as_u64_lenient,
-    get_numeric_or_date_column_types,
+    get_all_value_sources, get_dynamic_columns, get_missing_val_as_u64_lenient,
+    get_numeric_or_date_column_types, get_value_source,
 };
 use crate::aggregation::agg_req::{Aggregation, AggregationVariants, Aggregations};
 use crate::aggregation::bucket::{
@@ -22,14 +22,16 @@ use crate::aggregation::bucket::{
 use crate::aggregation::metric::{
     build_segment_stats_collector, AverageAggregation, CardinalityAggReqData,
     CardinalityAggregationReq, CountAggregation, ExtendedStatsAggregation, MaxAggregation,
-    MetricAggReqData, MinAggregation, SegmentCardinalityCollector, SegmentExtendedStatsCollector,
-    SegmentPercentilesCollector, StatsAggregation, StatsType, SumAggregation, TermOrdSet,
-    TopHitsAggReqData, TopHitsSegmentCollector, BITSET_MAX_TERM_ORD,
+    MetricAggReqData, MinAggregation, SegmentExtendedStatsCollector, SegmentPercentilesCollector,
+    StatsAggregation, StatsType, SumAggregation, TopHitsAggReqData, TopHitsSegmentCollector,
 };
 use crate::aggregation::segment_agg_result::{
     GenericSegmentAggregationResultsCollector, SegmentAggregationCollector,
 };
-use crate::aggregation::{f64_to_fastfield_u64, AggContextParams, ColumnBlockAccessor, Key};
+use crate::aggregation::{
+    f64_to_fastfield_u64, AggContextParams, ColumnBlockAccessor, Key, ValueSource,
+    ValueSourceRegistry,
+};
 use crate::{SegmentOrdinal, SegmentReader};
 
 #[derive(Default)]
@@ -37,8 +39,8 @@ use crate::{SegmentOrdinal, SegmentReader};
 /// It is passed to the collectors during collection.
 pub struct AggregationsSegmentCtx {
     /// Request data for each aggregation type.
-    pub per_request: PerRequestAggSegCtx,
-    pub context: AggContextParams,
+    pub(crate) per_request: PerRequestAggSegCtx,
+    pub(crate) context: AggContextParams,
     pub(crate) column_block_accessor: ColumnBlockAccessor,
 }
 
@@ -115,28 +117,28 @@ impl AggregationsSegmentCtx {
 #[derive(Default)]
 pub struct PerRequestAggSegCtx {
     /// TermsAggReqData contains the request data for a terms aggregation.
-    pub term_req_data: Vec<TermsAggReqData>,
+    pub(crate) term_req_data: Vec<TermsAggReqData>,
     /// HistogramAggReqData contains the request data for a histogram aggregation.
-    pub histogram_req_data: Vec<HistogramAggReqData>,
+    pub(crate) histogram_req_data: Vec<HistogramAggReqData>,
     /// RangeAggReqData contains the request data for a range aggregation.
-    pub range_req_data: Vec<RangeAggReqData>,
+    pub(crate) range_req_data: Vec<RangeAggReqData>,
     /// FilterAggReqData contains the request data for a filter aggregation.
-    pub filter_req_data: Vec<FilterAggReqData>,
+    pub(crate) filter_req_data: Vec<FilterAggReqData>,
     /// Shared by avg, min, max, sum, stats, extended_stats, count
-    pub stats_metric_req_data: Vec<MetricAggReqData>,
+    pub(crate) stats_metric_req_data: Vec<MetricAggReqData>,
     /// CardinalityAggReqData contains the request data for a cardinality aggregation.
-    pub cardinality_req_data: Vec<CardinalityAggReqData>,
+    pub(crate) cardinality_req_data: Vec<CardinalityAggReqData>,
     /// TopHitsAggReqData contains the request data for a top_hits aggregation.
-    pub top_hits_req_data: Vec<TopHitsAggReqData>,
+    pub(crate) top_hits_req_data: Vec<TopHitsAggReqData>,
     /// MissingTermAggReqData contains the request data for a missing term aggregation.
-    pub missing_term_req_data: Vec<MissingTermAggReqData>,
+    pub(crate) missing_term_req_data: Vec<MissingTermAggReqData>,
     /// CompositeAggReqData contains the request data for a composite aggregation.
-    pub composite_req_data: Vec<CompositeAggReqData>,
+    pub(crate) composite_req_data: Vec<CompositeAggReqData>,
     /// MultiTermsAggReqData contains the request data for a multi_terms aggregation.
-    pub multi_terms_req_data: Vec<MultiTermsAggReqData>,
+    pub(crate) multi_terms_req_data: Vec<MultiTermsAggReqData>,
 
     /// Request tree used to build collectors.
-    pub agg_tree: Vec<AggRefNode>,
+    pub(crate) agg_tree: Vec<AggRefNode>,
 }
 
 impl PerRequestAggSegCtx {
@@ -286,39 +288,7 @@ pub(crate) fn build_segment_agg_collector(
             Ok(Box::new(TermMissingAgg::new(req, node)?))
         }
         AggKind::Cardinality => {
-            let req_data = req.get_cardinality_req_data(node.idx_in_req_data);
-            // For str columns, choose the per-bucket entries representation
-            // based on the segment's column.max_value():
-            //   * small (< BITSET_MAX_TERM_ORD): `BitSet`, pre-allocated, no promotion machinery.
-            //   * large: `TermOrdSet` (sparse FxHashSet that promotes to a paged bitset).
-            // For non-str columns the `entries` field is unused (values go
-            // straight into the HLL sketch); we still pick `TermOrdSet`
-            // because its empty Sparse(FxHashSet) costs nothing.
-            let is_str = req_data.column_type == ColumnType::Str;
-            let max_term_ord_inclusive = if is_str {
-                req_data.accessor.max_value()
-            } else {
-                0
-            };
-            let collector: Box<dyn SegmentAggregationCollector> =
-                if is_str && max_term_ord_inclusive < BITSET_MAX_TERM_ORD {
-                    Box::new(SegmentCardinalityCollector::<BitSet>::from_req(
-                        req_data.column_type,
-                        node.idx_in_req_data,
-                        req_data.accessor.clone(),
-                        req_data.missing_value_for_accessor,
-                        max_term_ord_inclusive,
-                    ))
-                } else {
-                    Box::new(SegmentCardinalityCollector::<TermOrdSet>::from_req(
-                        req_data.column_type,
-                        node.idx_in_req_data,
-                        req_data.accessor.clone(),
-                        req_data.missing_value_for_accessor,
-                        max_term_ord_inclusive,
-                    ))
-                };
-            Ok(collector)
+            crate::aggregation::metric::build_segment_cardinality_collector(req, node)
         }
         AggKind::StatsKind(stats_type) => {
             let req_data = &mut req.per_request.stats_metric_req_data[node.idx_in_req_data];
@@ -336,7 +306,6 @@ pub(crate) fn build_segment_agg_collector(
                     let req_data = req.get_metric_req_data(node.idx_in_req_data);
                     Ok(Box::new(
                         SegmentPercentilesCollector::from_req_and_validate(
-                            req_data.field_type,
                             req_data.missing_u64,
                             req_data.accessor.clone(),
                             node.idx_in_req_data,
@@ -436,6 +405,46 @@ pub(crate) fn build_aggregations_data_from_req(
     Ok(data)
 }
 
+/// Resolves the substitute value used for documents that have none.
+///
+/// Only the `Str` arms of [`get_missing_val_as_u64_lenient`] read the column's max value — they
+/// place the sentinel one past the last real term ordinal — and that bound exists only for a
+/// materialized column. A computed text source therefore cannot support `missing`; for numeric
+/// types the argument is ignored, so any value will do.
+fn missing_value_for_source(
+    accessor: &dyn ValueSource,
+    missing: &Key,
+    field_name: &str,
+) -> crate::Result<Option<u64>> {
+    let column_type = accessor.column_type();
+    let column_max_value = match accessor.as_column() {
+        Some(column) => column.max_value(),
+        None if column_type == ColumnType::Str => {
+            return Err(crate::TantivyError::InvalidArgument(format!(
+                "`missing` is not supported for the computed text value source `{field_name}`"
+            )));
+        }
+        None => 0,
+    };
+    get_missing_val_as_u64_lenient(column_type, column_max_value, missing, field_name)
+}
+
+/// Extracts the materialized column, rejecting a computed source.
+///
+/// For the aggregations that read values through per-document random access or
+/// `ColumnIndex::has_value`, neither of which `ValueSource` can express.
+fn require_physical_column(
+    source: &dyn ValueSource,
+    field_name: &str,
+    agg_kind: &str,
+) -> crate::Result<Column<u64>> {
+    source.as_column().cloned().ok_or_else(|| {
+        crate::TantivyError::InvalidArgument(format!(
+            "{agg_kind} does not support the computed value source `{field_name}`"
+        ))
+    })
+}
+
 fn build_nodes(
     agg_name: &str,
     req: &Aggregation,
@@ -445,16 +454,17 @@ fn build_nodes(
     is_top_level: bool,
 ) -> crate::Result<Vec<AggRefNode>> {
     use AggregationVariants::*;
+    let value_sources = &data.context.value_sources;
     match &req.agg {
         Range(range_req) => {
-            let (accessor, field_type) = get_ff_reader(
+            let accessor = get_value_source(
                 reader,
+                value_sources,
                 &range_req.field,
                 Some(get_numeric_or_date_column_types()),
             )?;
             let idx_in_req_data = data.push_range_req_data(RangeAggReqData {
                 accessor,
-                field_type,
                 name: agg_name.to_string(),
                 req: range_req.clone(),
                 is_top_level,
@@ -467,14 +477,14 @@ fn build_nodes(
             }])
         }
         Histogram(histo_req) => {
-            let (accessor, field_type) = get_ff_reader(
+            let accessor = get_value_source(
                 reader,
+                value_sources,
                 &histo_req.field,
                 Some(get_numeric_or_date_column_types()),
             )?;
             let idx_in_req_data = data.push_histogram_req_data(HistogramAggReqData {
                 accessor,
-                field_type,
                 name: agg_name.to_string(),
                 req: histo_req.clone(),
                 is_date_histogram: false,
@@ -492,14 +502,17 @@ fn build_nodes(
             }])
         }
         DateHistogram(date_req) => {
-            let (accessor, field_type) =
-                get_ff_reader(reader, &date_req.field, Some(&[ColumnType::DateTime]))?;
+            let accessor = get_value_source(
+                reader,
+                value_sources,
+                &date_req.field,
+                Some(&[ColumnType::DateTime]),
+            )?;
             // Convert to histogram request, normalize to ns precision
             let mut histo_req = date_req.to_histogram_req()?;
             histo_req.normalize_date_time();
             let idx_in_req_data = data.push_histogram_req_data(HistogramAggReqData {
                 accessor,
-                field_type,
                 name: agg_name.to_string(),
                 req: histo_req,
                 is_date_histogram: true,
@@ -576,10 +589,10 @@ fn build_nodes(
                     ))
                 }
             };
-            let (accessor, field_type) = get_ff_reader(reader, field, allowed_column_types)?;
+            let accessor = get_value_source(reader, value_sources, field, allowed_column_types)?;
+            let field_type = accessor.column_type();
             let idx_in_req_data = data.push_metric_req_data(MetricAggReqData {
                 accessor,
-                field_type,
                 name: agg_name.to_string(),
                 collecting_for,
                 missing: *missing,
@@ -599,14 +612,15 @@ fn build_nodes(
         // Percentiles handled as Metric as well
         AggregationVariants::Percentiles(percentiles_req) => {
             percentiles_req.validate()?;
-            let (accessor, field_type) = get_ff_reader(
+            let accessor = get_value_source(
                 reader,
+                value_sources,
                 percentiles_req.field_name(),
                 Some(get_numeric_or_date_column_types()),
             )?;
+            let field_type = accessor.column_type();
             let idx_in_req_data = data.push_metric_req_data(MetricAggReqData {
                 accessor,
-                field_type,
                 name: agg_name.to_string(),
                 collecting_for: StatsType::Percentiles,
                 missing: percentiles_req.missing,
@@ -631,7 +645,18 @@ fn build_nodes(
             let accessors: Vec<(Column<u64>, ColumnType)> = top_hits
                 .field_names()
                 .iter()
-                .map(|field| get_ff_reader(reader, field, Some(get_numeric_or_date_column_types())))
+                .map(|field| {
+                    let source = get_value_source(
+                        reader,
+                        value_sources,
+                        field,
+                        Some(get_numeric_or_date_column_types()),
+                    )?;
+                    // Sort fields are read one document at a time via `values_for_doc`, which has
+                    // no block equivalent.
+                    let column = require_physical_column(&*source, field, "top_hits")?;
+                    Ok((column, source.column_type()))
+                })
                 .collect::<crate::Result<_>>()?;
 
             let value_accessors = top_hits
@@ -690,7 +715,6 @@ fn build_nodes(
 
             let idx_in_req_data = data.push_filter_req_data(FilterAggReqData {
                 name: agg_name.to_string(),
-                req: filter_req.clone(),
                 segment_reader: reader.clone(),
                 evaluator,
                 is_top_level,
@@ -749,11 +773,21 @@ fn build_multi_terms_nodes(
         ));
     }
 
+    let value_sources = data.context.value_sources.clone();
     let mut accessors_by_field = Vec::with_capacity(req.terms.len());
     for field_def in &req.terms {
         let field_name = &field_def.field;
         let str_dict_column = reader.fast_fields().str(field_name)?;
-        let columns = get_term_agg_accessors(reader, field_name, &field_def.missing, true)?;
+        // multi_terms resolves missing values through `ColumnIndex::has_value` per document, and
+        // exposes its columns on a public struct, so it stays physical-only.
+        let columns =
+            get_term_agg_accessors(reader, &value_sources, field_name, &field_def.missing, true)?
+                .into_iter()
+                .map(|source| {
+                    let column = require_physical_column(&*source, field_name, "multi_terms")?;
+                    Ok((column, source.column_type()))
+                })
+                .collect::<crate::Result<Vec<_>>>()?;
 
         if let Some((_, column_type)) = columns
             .iter()
@@ -827,7 +861,6 @@ fn build_multi_terms_nodes(
             req: req.clone(),
             fields,
             missing_accessors,
-            sub_aggregations: sub_aggs.clone(),
             is_top_level,
         });
         let children = build_children(sub_aggs, reader, segment_ordinal, data)?;
@@ -968,10 +1001,11 @@ fn build_children(
 
 fn get_term_agg_accessors(
     reader: &SegmentReader,
+    value_sources: &ValueSourceRegistry,
     field_name: &str,
     missing: &Option<Key>,
     include_bytes: bool,
-) -> crate::Result<Vec<(Column<u64>, ColumnType)>> {
+) -> crate::Result<Vec<Arc<dyn ValueSource>>> {
     // `terms` and `multi_terms` both explicitly reject `Bytes` columns downstream, which needs
     // to actually see them as a real column (rather than the empty shim below) to do so.
     // `cardinality` has no such rejection: it would hash raw `Bytes` term ordinals as if they
@@ -1002,14 +1036,15 @@ fn get_term_agg_accessors(
         })
         .unwrap_or(ColumnType::U64);
 
-    let column_and_types = get_all_ff_reader_or_empty(
+    let sources = get_all_value_sources(
         reader,
+        value_sources,
         field_name,
         Some(&allowed_column_types),
         fallback_type,
     )?;
 
-    Ok(column_and_types)
+    Ok(sources)
 }
 
 enum TermsOrCardinalityRequest {
@@ -1040,14 +1075,16 @@ fn build_terms_or_cardinality_nodes(
     let mut nodes = Vec::new();
 
     let str_dict_column = reader.fast_fields().str(field_name)?;
+    let value_sources = data.context.value_sources.clone();
 
     let include_bytes = matches!(req, TermsOrCardinalityRequest::Terms(_));
-    let column_and_types = get_term_agg_accessors(reader, field_name, missing, include_bytes)?;
+    let sources =
+        get_term_agg_accessors(reader, &value_sources, field_name, missing, include_bytes)?;
 
     // Special handling when missing + multi column or incompatible type on text/date.
-    let missing_and_more_than_one_col = column_and_types.len() > 1 && missing.is_some();
-    let text_on_non_text_col = column_and_types.len() == 1
-        && column_and_types[0].1 != ColumnType::Str
+    let missing_and_more_than_one_col = sources.len() > 1 && missing.is_some();
+    let text_on_non_text_col = sources.len() == 1
+        && sources[0].column_type() != ColumnType::Str
         && matches!(missing, Some(Key::Str(_)));
 
     let use_special_missing_agg = missing_and_more_than_one_col || text_on_non_text_col;
@@ -1064,9 +1101,21 @@ fn build_terms_or_cardinality_nodes(
                 Key::U64(_) => ColumnType::U64,
             })
             .unwrap_or(ColumnType::U64);
-        let all_accessors = get_all_ff_reader_or_empty(reader, field_name, None, fallback_type)?
-            .into_iter()
-            .collect::<Vec<_>>();
+        // This path inspects `ColumnIndex::has_value` per document per accessor to decide which
+        // documents are missing across several typed columns. There is no way to ask that through
+        // `ValueSource`, so it stays physical-only.
+        let all_accessors =
+            get_all_value_sources(reader, &value_sources, field_name, None, fallback_type)?
+                .into_iter()
+                .map(|source| {
+                    let column = require_physical_column(
+                        &*source,
+                        field_name,
+                        "terms with `missing` across multiple column types",
+                    )?;
+                    Ok((column, source.column_type()))
+                })
+                .collect::<crate::Result<Vec<_>>>()?;
         // This case only happens when we have term aggregation, or we fail
         let req = req.as_terms().cloned().ok_or_else(|| {
             crate::TantivyError::InvalidArgument(
@@ -1089,11 +1138,12 @@ fn build_terms_or_cardinality_nodes(
     }
 
     // Add one node per accessor
-    for (accessor, column_type) in column_and_types {
+    for accessor in sources {
+        let column_type = accessor.column_type();
         let missing_value_for_accessor = if use_special_missing_agg {
             None
         } else if let Some(m) = missing.as_ref() {
-            get_missing_val_as_u64_lenient(column_type, accessor.max_value(), m, field_name)?
+            missing_value_for_source(&*accessor, m, field_name)?
         } else {
             None
         };
@@ -1120,12 +1170,11 @@ fn build_terms_or_cardinality_nodes(
                 };
                 let idx_in_req_data = data.push_term_req_data(TermsAggReqData {
                     accessor,
-                    column_type,
                     str_dict_column: str_dict_column.clone(),
                     missing_value_for_accessor,
                     name: agg_name.to_string(),
                     req: TermsAggregationInternal::from_req(req),
-                    sug_aggregations: sub_aggs.clone(),
+                    sub_aggregations: sub_aggs.clone(),
                     allowed_term_ids,
                     is_top_level,
                 });
@@ -1144,7 +1193,6 @@ fn build_terms_or_cardinality_nodes(
                 };
                 let idx_in_req_data = data.push_cardinality_req_data(CardinalityAggReqData {
                     accessor,
-                    column_type,
                     str_dict_column: str_dict_column_for_req,
                     missing_value_for_accessor,
                     name: agg_name.to_string(),
diff --git a/src/aggregation/bucket/composite/accessors.rs b/src/aggregation/bucket/composite/accessors.rs
index 005700bed..c21e3a065 100644
--- a/src/aggregation/bucket/composite/accessors.rs
+++ b/src/aggregation/bucket/composite/accessors.rs
@@ -17,13 +17,13 @@ use crate::{SegmentReader, TantivyError};
 /// Contains all information required by the SegmentCompositeCollector to perform the
 /// composite aggregation on a segment.
 #[derive(Debug, Clone)]
-pub struct CompositeAggReqData {
+pub(crate) struct CompositeAggReqData {
     /// The name of the aggregation.
-    pub name: String,
+    pub(crate) name: String,
     /// The normalized term aggregation request.
-    pub req: CompositeAggregation,
+    pub(crate) req: CompositeAggregation,
     /// Accessors for each source, each source can have multiple accessors (columns).
-    pub composite_accessors: Vec<CompositeSourceAccessors>,
+    pub(crate) composite_accessors: Vec<CompositeSourceAccessors>,
 }
 
 impl CompositeAggReqData {
diff --git a/src/aggregation/bucket/composite/mod.rs b/src/aggregation/bucket/composite/mod.rs
index 1ea39f8a3..b7e5dd70a 100644
--- a/src/aggregation/bucket/composite/mod.rs
+++ b/src/aggregation/bucket/composite/mod.rs
@@ -15,8 +15,9 @@ use rustc_hash::FxHashMap;
 use serde::{Deserialize, Serialize};
 
 use crate::aggregation::agg_result::CompositeKey;
+pub(crate) use crate::aggregation::bucket::composite::accessors::CompositeAggReqData;
 pub use crate::aggregation::bucket::composite::accessors::{
-    CompositeAccessor, CompositeAggReqData, CompositeSourceAccessors, PrecomputedDateInterval,
+    CompositeAccessor, CompositeSourceAccessors, PrecomputedDateInterval,
 };
 pub use crate::aggregation::bucket::composite::collector::SegmentCompositeCollector;
 use crate::aggregation::bucket::composite::numeric_types::num_cmp::{
diff --git a/src/aggregation/bucket/filter.rs b/src/aggregation/bucket/filter.rs
index 9e39edd41..0c08d08c7 100644
--- a/src/aggregation/bucket/filter.rs
+++ b/src/aggregation/bucket/filter.rs
@@ -398,19 +398,17 @@ impl PartialEq for FilterAggregation {
 /// Request data for filter aggregation
 /// This struct holds the per-segment data needed to execute a filter aggregation
 #[derive(Clone)]
-pub struct FilterAggReqData {
+pub(crate) struct FilterAggReqData {
     /// The name of the filter aggregation
-    pub name: String,
-    /// The filter aggregation
-    pub req: FilterAggregation,
+    pub(crate) name: String,
     /// The segment reader
-    pub segment_reader: SegmentReader,
+    pub(crate) segment_reader: SegmentReader,
     /// Document evaluator for the filter query (precomputed BitSet).
     /// Wrapped in `Rc` so cloning the request data does not duplicate the (potentially large)
     /// underlying BitSet.
-    pub evaluator: Rc<DocumentQueryEvaluator>,
+    pub(crate) evaluator: Rc<DocumentQueryEvaluator>,
     /// True if this filter aggregation is at the top level of the aggregation tree (not nested).
-    pub is_top_level: bool,
+    pub(crate) is_top_level: bool,
 }
 
 impl FilterAggReqData {
diff --git a/src/aggregation/bucket/histogram/histogram.rs b/src/aggregation/bucket/histogram/histogram.rs
index e97a9ca5c..80079503d 100644
--- a/src/aggregation/bucket/histogram/histogram.rs
+++ b/src/aggregation/bucket/histogram/histogram.rs
@@ -1,6 +1,7 @@
 use std::cmp::Ordering;
+use std::sync::Arc;
 
-use columnar::{Column, ColumnType};
+use columnar::ColumnType;
 use rustc_hash::FxHashMap;
 use serde::{Deserialize, Serialize};
 use tantivy_bitpacker::minmax;
@@ -22,21 +23,19 @@ use crate::TantivyError;
 /// Contains all information required by the SegmentHistogramCollector to perform the
 /// histogram or date_histogram aggregation on a segment.
 #[derive(Debug, Clone)]
-pub struct HistogramAggReqData {
+pub(crate) struct HistogramAggReqData {
     /// The column accessor to access the fast field values.
-    pub accessor: Column<u64>,
-    /// The field type of the fast field.
-    pub field_type: ColumnType,
+    pub(crate) accessor: Arc<dyn ValueSource>,
     /// The name of the aggregation.
-    pub name: String,
+    pub(crate) name: String,
     /// The histogram aggregation request.
-    pub req: HistogramAggregation,
+    pub(crate) req: HistogramAggregation,
     /// True if this is a date_histogram aggregation.
-    pub is_date_histogram: bool,
+    pub(crate) is_date_histogram: bool,
     /// The bounds to limit the buckets to.
-    pub bounds: HistogramBounds,
+    pub(crate) bounds: HistogramBounds,
     /// The offset used to calculate the bucket position.
-    pub offset: f64,
+    pub(crate) offset: f64,
 }
 impl HistogramAggReqData {
     /// Estimate the memory consumption of this struct in bytes.
@@ -447,6 +446,7 @@ pub struct SegmentHistogramCollector<B> {
     parent_buckets: Vec<HistogramBuckets<B>>,
     sub_agg: Option<HighCardBufferedSubAggs>,
     req_data: HistogramAggReqData,
+    column_type: ColumnType,
     bucket_id_provider: BucketIdProvider,
     /// Theoretical bucket range derived from the column min/max, if dense `Vec` storage is
     /// viable. `None` keeps every parent bucket in the sparse hash map.
@@ -491,11 +491,11 @@ impl<B: BucketIdSlot> SegmentAggregationCollector for SegmentHistogramCollector<
 
         agg_data
             .column_block_accessor
-            .fetch_block(docs, &req.accessor);
+            .fetch_block(docs, &*req.accessor);
         // special path for nested buckets
         if let Some(sub_agg) = &mut self.sub_agg {
             for (doc, val) in agg_data.column_block_accessor.iter_docid_vals(docs) {
-                let val = f64_from_fastfield_u64(val, req.field_type);
+                let val = f64_from_fastfield_u64(val, self.column_type);
                 if bounds.contains(val) {
                     let bucket = store.get_or_create(
                         get_bucket_pos(val),
@@ -508,7 +508,7 @@ impl<B: BucketIdSlot> SegmentAggregationCollector for SegmentHistogramCollector<
             }
         } else {
             for val in agg_data.column_block_accessor.iter_vals() {
-                let val = f64_from_fastfield_u64(val, req.field_type);
+                let val = f64_from_fastfield_u64(val, self.column_type);
                 if bounds.contains(val) {
                     let bucket = store.get_or_create(
                         get_bucket_pos(val),
@@ -586,7 +586,7 @@ impl<B: BucketIdSlot> SegmentHistogramCollector<B> {
         }
         buckets.sort_unstable_by(|b1, b2| b1.key.total_cmp(&b2.key));
 
-        let is_date_agg = self.req_data.field_type == ColumnType::DateTime;
+        let is_date_agg = self.req_data.accessor.column_type() == ColumnType::DateTime;
         Ok(IntermediateBucketResult::Histogram {
             buckets,
             is_date_agg,
@@ -609,18 +609,18 @@ impl<B: BucketIdSlot> SegmentHistogramCollector<B> {
             .limits
             .add_memory_consumed(req_data.get_memory_consumption() as u64)?;
         let dense_range = compute_dense_range(
-            &req_data.accessor,
-            req_data.field_type,
+            &*req_data.accessor,
             req_data.req.interval,
             req_data.offset,
             req_data.bounds,
         );
         let sub_agg = sub_agg.map(BufferedSubAggs::new);
-
+        let column_type = req_data.accessor.column_type();
         Ok(Self {
             parent_buckets: Default::default(),
             sub_agg,
             req_data,
+            column_type,
             bucket_id_provider: BucketIdProvider::default(),
             dense_range,
         })
@@ -661,10 +661,12 @@ impl SegmentHistogramCollector<()> {
                 HistogramBuckets::Dense { base_pos, buckets }
             })
             .collect();
+        let column_type = req_data.accessor.column_type();
         Self {
             parent_buckets,
             sub_agg: None,
             req_data,
+            column_type,
             bucket_id_provider: BucketIdProvider::default(),
             dense_range: None,
         }
@@ -675,7 +677,8 @@ impl SegmentHistogramCollector<()> {
 /// `histogram` on a date column) and resolves `bounds`/`offset` from the request.
 fn normalize_histogram_req(req_data: &mut HistogramAggReqData) -> crate::Result<()> {
     req_data.req.validate()?;
-    if req_data.field_type == ColumnType::DateTime && !req_data.is_date_histogram {
+    let field_type = req_data.accessor.column_type();
+    if field_type == ColumnType::DateTime && !req_data.is_date_histogram {
         req_data.req.normalize_date_time();
     }
     req_data.bounds = req_data.req.hard_bounds.unwrap_or(HistogramBounds {
@@ -690,13 +693,15 @@ fn normalize_histogram_req(req_data: &mut HistogramAggReqData) -> crate::Result<
     // emission reads `req.hard_bounds` directly (see `get_req_min_max`), and `hard_bounds` only
     // ever clips that range, so a wider-than-data bound leaves the result unchanged.
     if req_data.req.hard_bounds.is_some() {
-        let col_min = f64_from_fastfield_u64(req_data.accessor.min_value(), req_data.field_type);
-        let col_max = f64_from_fastfield_u64(req_data.accessor.max_value(), req_data.field_type);
-        if col_min >= req_data.bounds.min && col_max <= req_data.bounds.max {
-            req_data.bounds = HistogramBounds {
-                min: f64::MIN,
-                max: f64::MAX,
-            };
+        if let Some((min_value, max_value)) = req_data.accessor.bounds() {
+            let col_min = f64_from_fastfield_u64(min_value, field_type);
+            let col_max = f64_from_fastfield_u64(max_value, field_type);
+            if col_min >= req_data.bounds.min && col_max <= req_data.bounds.max {
+                req_data.bounds = HistogramBounds {
+                    min: f64::MIN,
+                    max: f64::MAX,
+                };
+            }
         }
     }
     Ok(())
@@ -712,8 +717,7 @@ pub(crate) fn prepare_histogram_dense_range(
     let mut req_data = agg_data.per_request.histogram_req_data[node.idx_in_req_data].clone();
     normalize_histogram_req(&mut req_data)?;
     let dense_range = compute_dense_range(
-        &req_data.accessor,
-        req_data.field_type,
+        &*req_data.accessor,
         req_data.req.interval,
         req_data.offset,
         req_data.bounds,
@@ -752,15 +756,19 @@ pub(crate) fn get_bucket_pos_f64(val: f64, interval: f64, offset: f64) -> f64 {
 ///
 /// The column min/max bound every value the collector can see, so a `Vec` sized to this range can
 /// be indexed by `bucket_pos - base_pos` without any out-of-bounds check on the hot path.
+///
+/// Returns `None` for a computed source: there is no global range to size the `Vec` from, so the
+/// histogram keeps its sparse map. The result is identical, just without the dense fast path.
 fn compute_dense_range(
-    accessor: &Column<u64>,
-    field_type: ColumnType,
+    accessor: &dyn ValueSource,
     interval: f64,
     offset: f64,
     bounds: HistogramBounds,
 ) -> Option<DenseRange> {
-    let col_min = f64_from_fastfield_u64(accessor.min_value(), field_type);
-    let col_max = f64_from_fastfield_u64(accessor.max_value(), field_type);
+    let (min_value, max_value) = accessor.bounds()?;
+    let field_type = accessor.column_type();
+    let col_min = f64_from_fastfield_u64(min_value, field_type);
+    let col_max = f64_from_fastfield_u64(max_value, field_type);
     let lo = col_min.max(bounds.min);
     let hi = col_max.min(bounds.max);
     if lo > hi {
diff --git a/src/aggregation/bucket/multi_terms/mod.rs b/src/aggregation/bucket/multi_terms/mod.rs
index 5b3af5f08..0b2d53c23 100644
--- a/src/aggregation/bucket/multi_terms/mod.rs
+++ b/src/aggregation/bucket/multi_terms/mod.rs
@@ -1,3 +1,4 @@
+use std::collections::hash_map::Entry;
 use std::fmt::Debug;
 use std::net::Ipv6Addr;
 use std::sync::Arc;
@@ -158,23 +159,21 @@ impl MultiTermsFieldAccessor {
 
 /// Per-request data bundle passed to the segment collector.
 #[derive(Debug, Clone)]
-pub struct MultiTermsAggReqData {
+pub(crate) struct MultiTermsAggReqData {
     /// Aggregation name used to look up this entry in the result tree.
-    pub name: String,
+    pub(crate) name: String,
     /// Original request (needed for final-result conversion).
-    pub req: MultiTermsAggregation,
+    pub(crate) req: MultiTermsAggregation,
     /// One typed accessor per field listed in `req.terms`.
-    pub fields: Vec<MultiTermsFieldAccessor>,
+    pub(crate) fields: Vec<MultiTermsFieldAccessor>,
     /// Missing-value handling corresponding to `fields`. Only the designated physical accessor
     /// choice for each requested field carries `Some`, preventing duplicate missing buckets when
     /// type-specific collectors are merged.
-    pub missing_accessors: Vec<Option<MultiTermsMissingAccessor>>,
-    /// Sub-aggregation descriptor (empty when no sub-aggs).
-    pub sub_aggregations: Aggregations,
+    pub(crate) missing_accessors: Vec<Option<MultiTermsMissingAccessor>>,
     /// True if this multi_terms aggregation is at the top level of the aggregation tree
     /// (not nested). Used to gate the Vec/Paged packed-key storage tiers, which assume a
     /// bounded number of parent buckets (mirrors [`TermsAggReqData::is_top_level`]).
-    pub is_top_level: bool,
+    pub(crate) is_top_level: bool,
 }
 
 impl MultiTermsAggReqData {
@@ -230,7 +229,7 @@ fn fetch_field_block(
     let missing_value = block_missing_value(missing);
     block_accessor.fetch_block_with_missing_unique_per_doc(
         docs,
-        &field.column,
+        &(&field.column, field.column_type),
         missing_value,
         true,
     );
@@ -252,11 +251,8 @@ trait MultiTermsPacking: Clone + Debug + 'static {
     fn push_full_values<I>(&self, keys: &mut [Self::PackingType], field_idx: usize, values: I)
     where I: IntoIterator<Item = u64>;
 
-    fn unpack(
-        &self,
-        key: &Self::PackingType,
-        req_data: &MultiTermsAggReqData,
-    ) -> crate::Result<Vec<IntermediateKey>>;
+    /// Extract one field's raw value, allowing dictionary lookups to be batched by field.
+    fn unpack_value(&self, key: &Self::PackingType, field_idx: usize) -> u64;
 }
 
 #[derive(Clone, Debug)]
@@ -288,18 +284,8 @@ impl MultiTermsPacking for U64ArrayKeyPacking {
         }
     }
 
-    fn unpack(
-        &self,
-        key: &Self::PackingType,
-        req_data: &MultiTermsAggReqData,
-    ) -> crate::Result<Vec<IntermediateKey>> {
-        key.iter()
-            .zip(req_data.fields.iter())
-            .zip(req_data.missing_accessors.iter())
-            .map(|((value, field_acc), missing)| {
-                resolve_key_value(*value, field_acc, missing.as_ref())
-            })
-            .collect()
+    fn unpack_value(&self, key: &Self::PackingType, field_idx: usize) -> u64 {
+        key[field_idx]
     }
 }
 
@@ -332,20 +318,10 @@ impl MultiTermsPacking for PackedU64KeyPacking {
         }
     }
 
-    fn unpack(
-        &self,
-        key: &Self::PackingType,
-        req_data: &MultiTermsAggReqData,
-    ) -> crate::Result<Vec<IntermediateKey>> {
-        self.packs
-            .iter()
-            .zip(req_data.fields.iter())
-            .zip(req_data.missing_accessors.iter())
-            .map(|((pack, field), missing)| {
-                let offset = key.checked_shr(pack.shift).unwrap_or(0) & pack.mask;
-                resolve_key_value(offset + pack.min_value, field, missing.as_ref())
-            })
-            .collect()
+    fn unpack_value(&self, key: &Self::PackingType, field_idx: usize) -> u64 {
+        let pack = self.packs[field_idx];
+        let offset = key.checked_shr(pack.shift).unwrap_or(0) & pack.mask;
+        offset + pack.min_value
     }
 }
 
@@ -735,8 +711,8 @@ where
         let mut result_entries: FxHashMap<Vec<IntermediateKey>, IntermediateTermBucketEntry> =
             FxHashMap::with_capacity_and_hasher(entries.len(), Default::default());
 
-        for entry in entries {
-            let intermediate_key = packing.unpack(&entry.key, req_data)?;
+        let keys = resolve_bucket_keys(packing, &entries, req_data)?;
+        for (entry, intermediate_key) in entries.into_iter().zip(keys) {
             let mut sub_aggregation_res = IntermediateAggregationResults::default();
             if let Some(sub_agg_collector) = sub_agg_collector.as_deref_mut() {
                 sub_agg_collector.add_intermediate_aggregation_result(
@@ -749,12 +725,12 @@ where
             // Distinct encoded keys can resolve to the same public key (for example a synthetic
             // missing string and a real date). Merge rather than overwriting either contribution.
             match result_entries.entry(intermediate_key) {
-                std::collections::hash_map::Entry::Occupied(mut occupied) => {
+                Entry::Occupied(mut occupied) => {
                     let existing = occupied.get_mut();
                     existing.doc_count += doc_count;
                     existing.sub_aggregation.merge_fruits(sub_aggregation_res)?;
                 }
-                std::collections::hash_map::Entry::Vacant(vacant) => {
+                Entry::Vacant(vacant) => {
                     vacant.insert(IntermediateTermBucketEntry {
                         doc_count,
                         sub_aggregation: sub_aggregation_res,
@@ -1097,6 +1073,65 @@ where
     }))
 }
 
+/// Resolve only the retained candidates, batching string ordinals by field. Sorted lookups
+/// decode each dictionary block once and reuse decoded terms for repeated ordinals.
+fn resolve_bucket_keys<P: MultiTermsPacking, B>(
+    packing: &P,
+    entries: &[MultiTermsBucketEntry<P::PackingType, B>],
+    req_data: &MultiTermsAggReqData,
+) -> crate::Result<Vec<Vec<IntermediateKey>>> {
+    let mut keys: Vec<_> = (0..entries.len())
+        .map(|_| Vec::with_capacity(req_data.fields.len()))
+        .collect();
+    for (field_idx, (field, missing)) in req_data
+        .fields
+        .iter()
+        .zip(&req_data.missing_accessors)
+        .enumerate()
+    {
+        let mut ords_and_positions = Vec::new();
+        for (position, entry) in entries.iter().enumerate() {
+            let value = packing.unpack_value(&entry.key, field_idx);
+            if field.column_type == ColumnType::Str
+                && !missing
+                    .as_ref()
+                    .map(|m| m.missing_value == value)
+                    .unwrap_or(false)
+            {
+                ords_and_positions.push((value, position));
+            } else {
+                keys[position].push(resolve_key_value(value, field, missing.as_ref())?);
+            }
+        }
+        if ords_and_positions.is_empty() {
+            continue;
+        }
+        ords_and_positions.sort_unstable();
+        let (ords, positions): (Vec<_>, Vec<_>) = ords_and_positions.into_iter().unzip();
+        let fallback_dict = Dictionary::empty();
+        let dictionary = field
+            .str_dict_column
+            .as_ref()
+            .map(|column| column.dictionary())
+            .unwrap_or(&fallback_dict);
+        let mut decoded_keys = Vec::with_capacity(ords.len());
+        let all_found = dictionary.sorted_ords_to_term_cb(&ords, |term| {
+            decoded_keys.push(IntermediateKey::Str(
+                String::from_utf8(term.to_vec()).expect("term dict returned non-UTF-8"),
+            ));
+        })?;
+        if !all_found {
+            return Err(TantivyError::InternalError(
+                "multi_terms string ordinal not found in dictionary".to_string(),
+            ));
+        }
+        for (position, key) in positions.into_iter().zip(decoded_keys) {
+            keys[position].push(key);
+        }
+    }
+    Ok(keys)
+}
+
 /// Resolve one raw fast-field value, recognizing the configured missing encoding first.
 ///
 /// When the encoding reuses a string term ordinal, both paths resolve identically by construction.
@@ -1332,86 +1367,99 @@ impl IntermediateMultiTermsBucketResult {
 
         let req = MultiTermsAggregationInternal::from_req(req);
 
-        let mut buckets: Vec<MultiTermsBucketEntry> = self
-            .entries
-            .into_iter()
-            .filter(|(_, e)| e.doc_count >= req.min_doc_count)
-            .map(|(key_vec, entry)| {
-                let key_as_string = key_vec
-                    .iter()
-                    .map(|k| match k {
-                        // Bool keys need special-casing: `Key` form is numeric (1/0),  but
-                        // `key_as_string` must still carry the "true"/"false" string form.
-                        IntermediateKey::Bool(b) => b.to_string(),
-                        other => Key::from(other.clone()).to_string(),
-                    })
-                    .collect::<Vec<_>>()
-                    .join("|");
-                let keys: Vec<Key> = key_vec.into_iter().map(Key::from).collect();
-                Ok(MultiTermsBucketEntry {
-                    key_as_string,
-                    key: keys,
-                    doc_count: entry.doc_count,
-                    sub_aggregation: entry
-                        .sub_aggregation
-                        .into_final_result_internal(sub_aggregation_req, limits)?,
-                })
-            })
-            .collect::<crate::Result<_>>()?;
+        let mut entries = Vec::with_capacity(self.entries.len());
+        entries.extend(
+            self.entries
+                .into_iter()
+                .filter(|(_, e)| e.doc_count >= req.min_doc_count),
+        );
 
-        // Sort by order.
+        // Select the final buckets before formatting keys or finalizing their sub-aggregations.
         match &req.order.target {
             OrderTarget::Count => {
                 if req.order.order == Order::Desc {
-                    buckets.sort_unstable_by_key(|b| std::cmp::Reverse(b.doc_count));
+                    entries.sort_unstable_by_key(|(_, entry)| std::cmp::Reverse(entry.doc_count));
                 } else {
-                    buckets.sort_unstable_by_key(|b| b.doc_count);
+                    entries.sort_unstable_by_key(|(_, entry)| entry.doc_count);
                 }
             }
             OrderTarget::Key => {
-                buckets.sort_by(|left, right| {
-                    let cmp = left
-                        .key
-                        .iter()
-                        .zip(right.key.iter())
-                        .find_map(|(l, r)| {
-                            let c = l.partial_cmp(r)?;
-                            if c != std::cmp::Ordering::Equal {
-                                Some(c)
-                            } else {
-                                None
-                            }
-                        })
-                        .unwrap_or(std::cmp::Ordering::Equal);
-                    if req.order.order == Order::Asc {
-                        cmp
-                    } else {
-                        cmp.reverse()
-                    }
-                });
+                // Final keys have different ordering from intermediate keys (e.g. IPs are
+                // strings and bools are u64s). Cache the conversion rather than cloning per
+                // comparison.
+                let key = |(key_vec, _): &(Vec<IntermediateKey>, IntermediateTermBucketEntry)| {
+                    key_vec.iter().cloned().map(Key::from).collect::<Vec<_>>()
+                };
+                if req.order.order == Order::Asc {
+                    entries.sort_by_cached_key(key);
+                } else {
+                    entries.sort_by_cached_key(|entry| std::cmp::Reverse(key(entry)));
+                }
             }
             OrderTarget::SubAggregation(name) => {
                 let (agg_name, agg_property) = get_agg_name_and_property(name);
-                let mut buckets_with_val = buckets
+                let mut entries_with_val = entries
                     .into_iter()
-                    .map(|bucket| {
-                        let val = bucket
+                    .map(|entry| {
+                        let sub_req = sub_aggregation_req.get(agg_name).ok_or_else(|| {
+                            TantivyError::InternalError(format!(
+                                "Can't find aggregation {agg_name:?} in sub-aggregations"
+                            ))
+                        })?;
+                        // Only finalize the ordering metric. Its final value can differ from
+                        // the intermediate value, e.g. an empty sum defaults to zero.
+                        let metric = entry
+                            .1
                             .sub_aggregation
+                            .aggs_res
+                            .get(agg_name)
+                            .cloned()
+                            .unwrap_or_else(|| {
+                                crate::aggregation::intermediate_agg_result::empty_from_req(sub_req)
+                            });
+                        let val = metric
+                            .into_final_result(sub_req, limits)?
                             .get_value_from_aggregation(agg_name, agg_property)?
                             .unwrap_or(f64::MIN);
-                        Ok((bucket, val))
+                        Ok((entry, val))
                     })
                     .collect::<crate::Result<Vec<_>>>()?;
-                buckets_with_val.sort_by(|(_, v1), (_, v2)| match req.order.order {
+                entries_with_val.sort_by(|(_, v1), (_, v2)| match req.order.order {
                     Order::Desc => v2.total_cmp(v1),
                     Order::Asc => v1.total_cmp(v2),
                 });
-                buckets = buckets_with_val.into_iter().map(|(b, _)| b).collect();
+                entries = entries_with_val
+                    .into_iter()
+                    .map(|(entry, _)| entry)
+                    .collect();
             }
         }
 
         let (_before_cutoff, sum_other_from_final) =
-            cut_off_buckets(&mut buckets, req.size as usize, None);
+            cut_off_buckets(&mut entries, req.size as usize, None);
+
+        let mut buckets = Vec::with_capacity(entries.len());
+        for (key_vec, entry) in entries {
+            let key_as_string = key_vec
+                .iter()
+                .map(|k| match k {
+                    // Bool keys need special-casing: `Key` form is numeric (1/0), but
+                    // `key_as_string` must still carry the "true"/"false" string form.
+                    IntermediateKey::Bool(b) => b.to_string(),
+                    other => Key::from(other.clone()).to_string(),
+                })
+                .collect::<Vec<_>>()
+                .join("|");
+            let keys: Vec<Key> = key_vec.into_iter().map(Key::from).collect();
+            buckets.push(MultiTermsBucketEntry {
+                key_as_string,
+                key: keys,
+                doc_count: entry.doc_count,
+                sub_aggregation: entry
+                    .sub_aggregation
+                    .into_final_result_internal(sub_aggregation_req, limits)?,
+            });
+        }
 
         let doc_count_error_upper_bound = if req.show_term_doc_count_error {
             Some(self.doc_count_error_upper_bound)
@@ -1553,6 +1601,63 @@ mod tests {
         Ok(())
     }
 
+    #[test]
+    fn test_multi_terms_final_key_order() -> crate::Result<()> {
+        let keys = [
+            IntermediateKey::IpAddr("::ffff:10.0.0.2".parse().unwrap()),
+            IntermediateKey::IpAddr("::ffff:10.0.0.10".parse().unwrap()),
+            IntermediateKey::Str("0".to_string()),
+            IntermediateKey::Bool(false),
+            IntermediateKey::I64(-1),
+            IntermediateKey::U64(2),
+            IntermediateKey::F64(1.5),
+        ];
+        let expected = ["0", "10.0.0.10", "10.0.0.2", "-1", "false", "2", "1.5"];
+        for order in ["asc", "desc"] {
+            let req = serde_json::from_value(json!({
+                "terms": [{"field": "value"}],
+                "size": 6,
+                "order": {"_key": order}
+            }))?;
+            let intermediate = IntermediateMultiTermsBucketResult {
+                entries: keys
+                    .iter()
+                    .cloned()
+                    .map(|key| {
+                        (
+                            vec![key],
+                            IntermediateTermBucketEntry {
+                                doc_count: 1,
+                                sub_aggregation: Default::default(),
+                            },
+                        )
+                    })
+                    .collect(),
+                ..Default::default()
+            };
+            let result = intermediate.into_final_result(
+                &req,
+                &Default::default(),
+                &mut Default::default(),
+            )?;
+            let result = serde_json::to_value(result)?;
+            let actual: Vec<_> = result["buckets"]
+                .as_array()
+                .unwrap()
+                .iter()
+                .map(|bucket| bucket["key_as_string"].as_str().unwrap())
+                .collect();
+            let mut expected = expected.to_vec();
+            if order == "desc" {
+                expected.reverse();
+            }
+            expected.truncate(6);
+            assert_eq!(actual, expected);
+            assert_eq!(result["sum_other_doc_count"], 1);
+        }
+        Ok(())
+    }
+
     #[test]
     fn test_multi_terms_min_doc_count() -> crate::Result<()> {
         let index = build_two_field_index(
@@ -1603,7 +1708,7 @@ mod tests {
         let buckets = res["mt"]["buckets"].as_array().unwrap();
         assert_eq!(buckets.len(), 2);
         assert_eq!(buckets[0]["key_as_string"], "rock|A");
-        assert!(res["mt"]["sum_other_doc_count"].as_u64().unwrap() > 0);
+        assert_eq!(res["mt"]["sum_other_doc_count"], 1);
         Ok(())
     }
 
@@ -2706,7 +2811,7 @@ mod tests {
     }
 
     #[test]
-    fn test_multi_terms_missing_subagg_value_sorts_last_at_segment_cutoff() -> crate::Result<()> {
+    fn test_multi_terms_missing_subagg_value_at_cutoff() -> crate::Result<()> {
         let mut schema_builder = Schema::builder();
         let genre_field = schema_builder.add_text_field("genre", STRING | FAST);
         let product_field = schema_builder.add_text_field("product", STRING | FAST);
@@ -2723,23 +2828,33 @@ mod tests {
             writer.commit()?;
         }
 
-        let agg_req: Aggregations = serde_json::from_value(json!({
-            "mt": {
-                "multi_terms": {
-                    "terms": [{"field": "genre"}, {"field": "product"}],
-                    "size": 1,
-                    "segment_size": 1,
-                    "order": {"avg_delta": "desc"}
-                },
-                "aggs": {
-                    "avg_delta": {"avg": {"field": "delta"}}
+        for (metric, segment_size, expected) in [
+            (json!({"avg": {"field": "delta"}}), 1, "rock|A"),
+            (json!({"avg": {"field": "delta"}}), 2, "rock|A"),
+            // At final cutoff, an empty sum is zero unless none_if_no_match is set.
+            (json!({"sum": {"field": "delta"}}), 2, "rock|B"),
+            (
+                json!({"sum": {"field": "delta", "none_if_no_match": true}}),
+                2,
+                "rock|A",
+            ),
+        ] {
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "mt": {
+                    "multi_terms": {
+                        "terms": [{"field": "genre"}, {"field": "product"}],
+                        "size": 1,
+                        "segment_size": segment_size,
+                        "order": {"metric": "desc"}
+                    },
+                    "aggs": {"metric": metric}
                 }
-            }
-        }))?;
-        let res = exec_request(agg_req, &index)?;
-        let buckets = res["mt"]["buckets"].as_array().unwrap();
-        assert_eq!(buckets.len(), 1);
-        assert_eq!(buckets[0]["key_as_string"], "rock|A");
+            }))?;
+            let res = exec_request(agg_req, &index)?;
+            let buckets = res["mt"]["buckets"].as_array().unwrap();
+            assert_eq!(buckets.len(), 1);
+            assert_eq!(buckets[0]["key_as_string"], expected);
+        }
         Ok(())
     }
 
diff --git a/src/aggregation/bucket/range.rs b/src/aggregation/bucket/range.rs
index a66ea8225..e2b581367 100644
--- a/src/aggregation/bucket/range.rs
+++ b/src/aggregation/bucket/range.rs
@@ -1,7 +1,8 @@
 use std::fmt::Debug;
 use std::ops::Range;
+use std::sync::Arc;
 
-use columnar::{Column, ColumnType};
+use columnar::ColumnType;
 use rustc_hash::FxHashMap;
 use serde::{Deserialize, Serialize};
 
@@ -18,23 +19,22 @@ use crate::aggregation::intermediate_agg_result::{
     IntermediateRangeBucketEntry, IntermediateRangeBucketResult,
 };
 use crate::aggregation::segment_agg_result::{BucketIdProvider, SegmentAggregationCollector};
+use crate::aggregation::value_source::ValueSource;
 use crate::aggregation::*;
 use crate::TantivyError;
 
 /// Contains all information required by the SegmentRangeCollector to perform the
 /// range aggregation on a segment.
 #[derive(Debug, Clone)]
-pub struct RangeAggReqData {
+pub(crate) struct RangeAggReqData {
     /// The column accessor to access the fast field values.
-    pub accessor: Column<u64>,
-    /// The type of the fast field.
-    pub field_type: ColumnType,
+    pub(crate) accessor: Arc<dyn ValueSource>,
     /// The range aggregation request.
-    pub req: RangeAggregation,
+    pub(crate) req: RangeAggregation,
     /// The name of the aggregation.
-    pub name: String,
+    pub(crate) name: String,
     /// Whether this is a top-level aggregation.
-    pub is_top_level: bool,
+    pub(crate) is_top_level: bool,
 }
 
 impl RangeAggReqData {
@@ -161,7 +161,6 @@ pub struct SegmentRangeCollector<B: SubAggBuffer> {
     /// The buckets containing the aggregation data.
     /// One for each ParentBucketId
     parent_buckets: Vec<Vec<SegmentRangeAndBucketEntry>>,
-    column_type: ColumnType,
     pub(crate) req_data: RangeAggReqData,
     sub_agg: Option<BufferedSubAggs<B>>,
     /// Here things get a bit weird. We need to assign unique bucket ids across all
@@ -184,7 +183,7 @@ impl<B: SubAggBuffer> Debug for SegmentRangeCollector<B> {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         f.debug_struct("SegmentRangeCollector")
             .field("parent_buckets_len", &self.parent_buckets.len())
-            .field("column_type", &self.column_type)
+            .field("column_type", &self.req_data.accessor.column_type())
             .field("name", &self.req_data.name)
             .field("has_sub_agg", &self.sub_agg.is_some())
             .finish()
@@ -239,7 +238,7 @@ impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentRangeCollector<B> {
         parent_bucket_id: BucketId,
     ) -> crate::Result<()> {
         self.prepare_max_bucket(parent_bucket_id, agg_data)?;
-        let field_type = self.column_type;
+        let field_type = self.req_data.accessor.column_type();
         let name = self.req_data.name.to_string();
 
         let buckets = std::mem::take(&mut self.parent_buckets[parent_bucket_id as usize]);
@@ -264,7 +263,7 @@ impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentRangeCollector<B> {
 
         let bucket = IntermediateBucketResult::Range(IntermediateRangeBucketResult {
             buckets,
-            column_type: Some(self.column_type),
+            column_type: Some(field_type),
         });
 
         results.push(name, IntermediateAggregationResult::Bucket(bucket))?;
@@ -281,7 +280,7 @@ impl<B: SubAggBuffer> SegmentAggregationCollector for SegmentRangeCollector<B> {
     ) -> crate::Result<()> {
         agg_data
             .column_block_accessor
-            .fetch_block(docs, &self.req_data.accessor);
+            .fetch_block(docs, &*self.req_data.accessor);
 
         let buckets = &mut self.parent_buckets[parent_bucket_id as usize];
 
@@ -343,7 +342,6 @@ pub(crate) fn build_segment_range_collector(
         .context
         .limits
         .add_memory_consumed(req_data.get_memory_consumption() as u64)?;
-    let field_type = req_data.field_type;
 
     // TODO: A better metric instead of is_top_level would be the number of buckets expected.
     // E.g. If range agg is not top level, but the parent is a bucket agg with less than 10 buckets,
@@ -359,7 +357,6 @@ pub(crate) fn build_segment_range_collector(
     if is_low_card {
         Ok(Box::new(SegmentRangeCollector::<LowCardSubAggBuffer> {
             sub_agg: sub_agg.map(LowCardBufferedSubAggs::new),
-            column_type: field_type,
             req_data,
             parent_buckets: Vec::new(),
             bucket_id_provider: BucketIdProvider::default(),
@@ -368,7 +365,6 @@ pub(crate) fn build_segment_range_collector(
     } else {
         Ok(Box::new(SegmentRangeCollector::<HighCardSubAggBuffer> {
             sub_agg: sub_agg.map(BufferedSubAggs::new),
-            column_type: field_type,
             req_data,
             parent_buckets: Vec::new(),
             bucket_id_provider: BucketIdProvider::default(),
@@ -379,8 +375,8 @@ pub(crate) fn build_segment_range_collector(
 
 impl<B: SubAggBuffer> SegmentRangeCollector<B> {
     pub(crate) fn create_new_buckets(&mut self) -> crate::Result<Vec<SegmentRangeAndBucketEntry>> {
-        let field_type = self.column_type;
         let req_data = &self.req_data;
+        let field_type = req_data.accessor.column_type();
         // The range input on the request is f64.
         // We need to convert to u64 ranges, because we read the values as u64.
         // The mapping from the conversion is monotonic so ordering is preserved.
diff --git a/src/aggregation/bucket/term_agg/flattened_term_histogram.rs b/src/aggregation/bucket/term_agg/flattened_term_histogram.rs
index 675171448..3f741539b 100644
--- a/src/aggregation/bucket/term_agg/flattened_term_histogram.rs
+++ b/src/aggregation/bucket/term_agg/flattened_term_histogram.rs
@@ -5,8 +5,9 @@
 //! [`maybe_build_flattened_collector`] for the conditions under which it is used.
 
 use std::fmt::Debug;
+use std::sync::Arc;
 
-use columnar::{Column, ColumnType};
+use columnar::{Cardinality, ColumnType, ColumnValues};
 
 use super::{
     Bucket, SegmentTermCollector, TermsAggReqData, VecTermBuckets, MAX_NUM_BUCKETS_FOR_COUNT_LANES,
@@ -14,7 +15,7 @@ use super::{
 };
 use crate::aggregation::agg_data::{AggKind, AggRefNode, AggregationsSegmentCtx};
 use crate::aggregation::bucket::{
-    get_bucket_pos_f64, prepare_histogram_dense_range, HistogramAggReqData,
+    get_bucket_pos_f64, prepare_histogram_dense_range, DenseRange, HistogramAggReqData,
     SegmentHistogramCollector,
 };
 use crate::aggregation::buffered_sub_aggs::LowCardSubAggBuffer;
@@ -22,7 +23,8 @@ use crate::aggregation::intermediate_agg_result::{
     IntermediateAggregationResult, IntermediateAggregationResults,
 };
 use crate::aggregation::segment_agg_result::{BucketIdProvider, SegmentAggregationCollector};
-use crate::aggregation::{f64_from_fastfield_u64, BucketId, ColumnBlockAccessor};
+use crate::aggregation::value_source::ColumnBlockAccessor;
+use crate::aggregation::{f64_from_fastfield_u64, BucketId, ValueSource};
 
 /// Maximum number of physical counters in the flattened flat grid. Above this the grid would be too
 /// large/cache-unfriendly, so we fall back to the general buffered path. Count lanes are included
@@ -39,7 +41,7 @@ const SINGLE_COUNT_LANE: usize = 1;
 const NUM_SMALL_LINEAR_BUCKETS: usize = 4;
 const NUM_LARGE_LINEAR_BUCKETS: usize = 8;
 
-trait BucketResolver: Debug + 'static {
+trait BucketResolver: 'static {
     /// Fetches the histogram values needed for this block. Resolvers that do not inspect the
     /// histogram column (notably [`FlattenedSingleBucketResolver`]) leave this as a no-op.
     fn prepare_block(&mut self, docs: &[crate::DocId]);
@@ -81,21 +83,11 @@ fn increment_grid_count<const LANES: usize>(
 
 /// Resolver for a histogram whose entire value range maps to one bucket. It deliberately owns no
 /// block accessor: collecting this shape does not read or decode the histogram column at all.
-#[derive(Debug)]
+#[derive(Default)]
 struct SingleBucketResolver {
     next_count_lane: usize,
 }
 
-impl SingleBucketResolver {
-    fn new(hist_req_data: &HistogramAggReqData) -> Self {
-        assert!(
-            hist_req_data.accessor.get_cardinality().is_full(),
-            "SingleBucketResolver requires a full histogram column"
-        );
-        Self { next_count_lane: 0 }
-    }
-}
-
 impl BucketResolver for SingleBucketResolver {
     #[inline]
     fn prepare_block(&mut self, _docs: &[crate::DocId]) {}
@@ -129,11 +121,10 @@ impl BucketResolver for SingleBucketResolver {
 
 /// The general resolver. It preserves the existing field conversion and floating-point bucket
 /// calculation for histograms that do not use a specialized resolver.
-#[derive(Debug)]
 struct ComputedBucketResolver {
     hist_block: ColumnBlockAccessor,
     next_count_lane: usize,
-    accessor: Column<u64>,
+    column_values: Arc<dyn ColumnValues>,
     field_type: ColumnType,
     interval: f64,
     offset: f64,
@@ -143,16 +134,17 @@ struct ComputedBucketResolver {
 }
 
 impl ComputedBucketResolver {
-    fn new(hist_req_data: &HistogramAggReqData, base_pos: i64, num_buckets: usize) -> Self {
-        assert!(
-            hist_req_data.accessor.get_cardinality().is_full(),
-            "ComputedBucketResolver requires a full histogram column"
-        );
+    fn new(
+        hist_req_data: &HistogramAggReqData,
+        hist_values: Arc<dyn ColumnValues>,
+        base_pos: i64,
+        num_buckets: usize,
+    ) -> Self {
         Self {
             hist_block: ColumnBlockAccessor::default(),
             next_count_lane: 0,
-            accessor: hist_req_data.accessor.clone(),
-            field_type: hist_req_data.field_type,
+            column_values: hist_values,
+            field_type: hist_req_data.accessor.column_type(),
             interval: hist_req_data.req.interval,
             offset: hist_req_data.offset,
             base_pos,
@@ -166,7 +158,7 @@ impl BucketResolver for ComputedBucketResolver {
     #[inline]
     fn prepare_block(&mut self, docs: &[crate::DocId]) {
         self.hist_block
-            .fetch_full_column_block(docs, &self.accessor);
+            .fetch_full_column_block(docs, &*self.column_values);
     }
 
     #[inline]
@@ -223,32 +215,30 @@ impl BucketResolver for ComputedBucketResolver {
 
 /// Resolver for a small histogram grid. Bucket starts are precomputed in monotonic fast-field
 /// `u64` space, then scanned linearly. `NUM_BUCKETS` is fixed so the optimizer can unroll the scan.
-#[derive(Debug)]
 struct LinearBucketResolver<const NUM_BUCKETS: usize> {
     hist_block: ColumnBlockAccessor,
     next_count_lane: usize,
-    accessor: Column<u64>,
+    column_values: Arc<dyn ColumnValues>,
     boundaries: [u64; NUM_BUCKETS],
     num_buckets: usize,
 }
 
 impl<const NUM_BUCKETS: usize> LinearBucketResolver<NUM_BUCKETS> {
+    /// Returns `None` only when no padding sentinel exists (see below), in which case the caller
+    /// falls back to [`ComputedBucketResolver`].
     fn new(
         hist_req_data: &HistogramAggReqData,
+        column_values: Arc<dyn ColumnValues>,
         base_pos: i64,
         num_time_buckets: usize,
     ) -> Option<Self> {
         assert!(num_time_buckets > 1 && num_time_buckets <= NUM_BUCKETS);
-        assert!(
-            hist_req_data.accessor.get_cardinality().is_full(),
-            "LinearBucketResolver requires a full histogram column"
-        );
-        let max_encoded_value = hist_req_data.accessor.max_value();
+        let max_encoded_value = column_values.max_value();
         // Padding must compare false for every column value. There is no such `u64` sentinel when
         // the column contains `u64::MAX`, so that edge case uses the computed resolver instead.
         let padding = max_encoded_value.checked_add(1)?;
         let mut boundaries = [padding; NUM_BUCKETS];
-        let mut bucket_start = hist_req_data.accessor.min_value();
+        let mut bucket_start = column_values.min_value();
         for bucket in 1..num_time_buckets {
             bucket_start = first_encoded_value_for_bucket(
                 bucket_start,
@@ -262,7 +252,7 @@ impl<const NUM_BUCKETS: usize> LinearBucketResolver<NUM_BUCKETS> {
         Some(Self {
             hist_block: ColumnBlockAccessor::default(),
             next_count_lane: 0,
-            accessor: hist_req_data.accessor.clone(),
+            column_values,
             boundaries,
             num_buckets: num_time_buckets,
         })
@@ -282,7 +272,7 @@ impl<const NUM_BUCKETS: usize> BucketResolver for LinearBucketResolver<NUM_BUCKE
     #[inline]
     fn prepare_block(&mut self, docs: &[crate::DocId]) {
         self.hist_block
-            .fetch_full_column_block(docs, &self.accessor);
+            .fetch_full_column_block(docs, &*self.column_values);
     }
 
     #[inline]
@@ -328,10 +318,11 @@ fn first_encoded_value_for_bucket(
     hist_req_data: &HistogramAggReqData,
     base_pos: i64,
 ) -> u64 {
+    let field_type = hist_req_data.accessor.column_type();
     while encoded_lower_bound < encoded_upper_bound {
         let encoded_midpoint =
             encoded_lower_bound + (encoded_upper_bound - encoded_lower_bound) / 2;
-        let val = f64_from_fastfield_u64(encoded_midpoint, hist_req_data.field_type);
+        let val = f64_from_fastfield_u64(encoded_midpoint, field_type);
         let bucket = (get_bucket_pos_f64(val, hist_req_data.req.interval, hist_req_data.offset)
             as i64
             - base_pos) as usize;
@@ -359,7 +350,6 @@ fn first_encoded_value_for_bucket(
 /// At result time the flat grid is expanded back into the regular term map + histogram storage and
 /// handed to the shared intermediate-result builders, so cross-segment merging is identical to the
 /// general path.
-#[derive(Debug)]
 struct FlattenedTermHistogramCollector<R: BucketResolver, const LANES: usize> {
     /// Per-term count of docs *outside* `hard_bounds` (still in `doc_count`, but in no bucket).
     /// Per-term total = this + the term's `counts` row-sum; left empty when there are no hard
@@ -375,7 +365,8 @@ struct FlattenedTermHistogramCollector<R: BucketResolver, const LANES: usize> {
     /// `bucket_pos` mapped to time-bucket index 0.
     base_pos: i64,
     terms_req_data: TermsAggReqData,
-    /// The (cloned, normalized) histogram request: its column + interval/offset/bounds.
+    /// The terms full column's values
+    terms_values: Arc<dyn ColumnValues>,
     hist_req_data: HistogramAggReqData,
     /// Private term block accessor. The bucket resolver owns a histogram block accessor when it
     /// needs one; the single-bucket resolver deliberately does not.
@@ -385,6 +376,15 @@ struct FlattenedTermHistogramCollector<R: BucketResolver, const LANES: usize> {
     all_docs_in_bounds: bool,
 }
 
+impl<R: BucketResolver, const LANES: usize> Debug for FlattenedTermHistogramCollector<R, LANES> {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        f.debug_struct("FlattenedTermHistogramCollector")
+            .field("base_pos", &self.base_pos)
+            .field("all_docs_in_bounds", &self.all_docs_in_bounds)
+            .finish_non_exhaustive()
+    }
+}
+
 impl<R: BucketResolver, const LANES: usize> SegmentAggregationCollector
     for FlattenedTermHistogramCollector<R, LANES>
 {
@@ -455,7 +455,7 @@ impl<R: BucketResolver, const LANES: usize> SegmentAggregationCollector
         // The term column is always needed. The resolver fetches the histogram column only when
         // bucket selection depends on its values; `SingleBucketResolver` makes this a no-op.
         self.term_block
-            .fetch_full_column_block(docs, &self.terms_req_data.accessor);
+            .fetch_full_column_block(docs, &*self.terms_values);
         self.bucket_resolver.prepare_block(docs);
 
         // Keep separate bounded and unbounded entry points so the common path has no bounds branch,
@@ -527,25 +527,26 @@ pub(super) fn maybe_build_flattened_collector(
     let fuseable = is_top_level
         // TODO: We can easily support this
         && terms_req_data.allowed_term_ids.is_none()
-        && terms_req_data.accessor.get_cardinality().is_full()
-        // The flat counters are `u32`, bumped once per value, so no count can exceed the column's
-        // value count. (Essentially always true here: the column is full, so its value count
-        // equals the doc count, and `DocId` is `u32`.)
-        && terms_req_data.accessor.values.num_vals() < u32::MAX
         && node.children.len() == 1
         && matches!(
             node.children[0].kind,
             AggKind::Histogram | AggKind::DateHistogram
         )
-        && node.children[0].children.is_empty()
-        && agg_data.per_request.histogram_req_data[node.children[0].idx_in_req_data]
-            .accessor
-            .get_cardinality()
-            .is_full();
+        && node.children[0].children.is_empty();
     if !fuseable {
         return Ok(None);
     }
 
+    // Check fullness once, here, and hand the bare values array downstream.
+    let Some(terms_values) = try_get_column_full_values(&*terms_req_data.accessor) else {
+        return Ok(None);
+    };
+    // The flat counters are `u32`, bumped once per value, so no count can exceed the
+    // column's value count. (Essentially always true here: the column is full, so its
+    // value count equals the doc count, and `DocId` is `u32`.)
+    if terms_values.num_vals() == u32::MAX {
+        return Ok(None);
+    }
     // Clone + normalize the histogram request and get its dense bucket range; only take the
     // flattened path when the physical counter grid is small enough. Very small logical grids use
     // multiple counters per cell; larger grids retain scalar cells to avoid paying for lanes when
@@ -554,6 +555,10 @@ pub(super) fn maybe_build_flattened_collector(
     else {
         return Ok(None);
     };
+    let Some(hist_values) = try_get_column_full_values(&*hist_req_data.accessor) else {
+        return Ok(None);
+    };
+
     let num_terms = col_max_val.saturating_add(1) as usize;
     let num_grid_cells = num_terms.saturating_mul(range.len);
     let use_count_lanes = num_grid_cells <= MAX_NUM_BUCKETS_FOR_COUNT_LANES;
@@ -571,40 +576,55 @@ pub(super) fn maybe_build_flattened_collector(
             agg_data,
             terms_req_data,
             hist_req_data,
+            terms_values,
+            hist_values,
             num_terms,
-            range.len,
-            range.base_pos,
+            range,
         )?
     } else {
         build_flattened_collector::<SINGLE_COUNT_LANE>(
             agg_data,
             terms_req_data,
             hist_req_data,
+            terms_values,
+            hist_values,
             num_terms,
-            range.len,
-            range.base_pos,
+            range,
         )?
     };
     Ok(Some(collector))
 }
 
+fn try_get_column_full_values(value_source: &dyn ValueSource) -> Option<Arc<dyn ColumnValues>> {
+    let column = value_source.as_column()?;
+    if column.get_cardinality() == Cardinality::Full {
+        Some(column.values.clone())
+    } else {
+        None
+    }
+}
+
 fn build_flattened_collector<const LANES: usize>(
     agg_data: &mut AggregationsSegmentCtx,
     terms_req_data: &TermsAggReqData,
     hist_req_data: HistogramAggReqData,
+    terms_values: Arc<dyn ColumnValues>,
+    hist_values: Arc<dyn ColumnValues>,
     num_terms: usize,
-    num_time_buckets: usize,
-    base_pos: i64,
+    range: DenseRange,
 ) -> crate::Result<Box<dyn SegmentAggregationCollector>> {
+    let num_time_buckets = range.len;
+    let base_pos = range.base_pos;
     const { assert!(LANES > 0, "a flattened grid needs at least one count lane") };
 
     let all_docs_in_bounds =
         hist_req_data.bounds.min == f64::MIN && hist_req_data.bounds.max == f64::MAX;
     if all_docs_in_bounds && num_time_buckets == 1 {
-        let resolver = SingleBucketResolver::new(&hist_req_data);
+        let resolver = SingleBucketResolver::default();
         return build_flattened_collector_with_resolver::<SingleBucketResolver, LANES>(
             agg_data,
             terms_req_data,
+            terms_values,
             hist_req_data,
             num_terms,
             base_pos,
@@ -614,12 +634,14 @@ fn build_flattened_collector<const LANES: usize>(
     if all_docs_in_bounds && num_time_buckets <= NUM_SMALL_LINEAR_BUCKETS {
         if let Some(resolver) = LinearBucketResolver::<NUM_SMALL_LINEAR_BUCKETS>::new(
             &hist_req_data,
+            hist_values.clone(),
             base_pos,
             num_time_buckets,
         ) {
             return build_flattened_collector_with_resolver::<_, LANES>(
                 agg_data,
                 terms_req_data,
+                terms_values,
                 hist_req_data,
                 num_terms,
                 base_pos,
@@ -629,12 +651,14 @@ fn build_flattened_collector<const LANES: usize>(
     } else if all_docs_in_bounds && num_time_buckets <= NUM_LARGE_LINEAR_BUCKETS {
         if let Some(resolver) = LinearBucketResolver::<NUM_LARGE_LINEAR_BUCKETS>::new(
             &hist_req_data,
+            hist_values.clone(),
             base_pos,
             num_time_buckets,
         ) {
             return build_flattened_collector_with_resolver::<_, LANES>(
                 agg_data,
                 terms_req_data,
+                terms_values,
                 hist_req_data,
                 num_terms,
                 base_pos,
@@ -643,10 +667,16 @@ fn build_flattened_collector<const LANES: usize>(
         }
     }
 
-    let resolver = ComputedBucketResolver::new(&hist_req_data, base_pos, num_time_buckets);
+    let resolver = ComputedBucketResolver::new(
+        &hist_req_data,
+        hist_values.clone(),
+        base_pos,
+        num_time_buckets,
+    );
     build_flattened_collector_with_resolver::<_, LANES>(
         agg_data,
         terms_req_data,
+        terms_values,
         hist_req_data,
         num_terms,
         base_pos,
@@ -657,6 +687,7 @@ fn build_flattened_collector<const LANES: usize>(
 fn build_flattened_collector_with_resolver<R: BucketResolver, const LANES: usize>(
     agg_data: &mut AggregationsSegmentCtx,
     terms_req_data: &TermsAggReqData,
+    terms_values: Arc<dyn ColumnValues<u64>>,
     hist_req_data: HistogramAggReqData,
     num_terms: usize,
     base_pos: i64,
@@ -686,6 +717,7 @@ fn build_flattened_collector_with_resolver<R: BucketResolver, const LANES: usize
         counts,
         base_pos,
         terms_req_data: terms_req_data.clone(),
+        terms_values,
         hist_req_data,
         term_block: ColumnBlockAccessor::default(),
         bucket_resolver,
diff --git a/src/aggregation/bucket/term_agg/mod.rs b/src/aggregation/bucket/term_agg/mod.rs
index 08ec2f093..c722d23f2 100644
--- a/src/aggregation/bucket/term_agg/mod.rs
+++ b/src/aggregation/bucket/term_agg/mod.rs
@@ -1,10 +1,11 @@
 use std::fmt::Debug;
 use std::hash::Hash;
 use std::net::Ipv6Addr;
+use std::sync::Arc;
 
 use columnar::column_values::CompactSpaceU64Accessor;
 use columnar::{
-    Column, ColumnType, Dictionary, MonotonicallyMappableToU128, MonotonicallyMappableToU64,
+    ColumnType, Dictionary, MonotonicallyMappableToU128, MonotonicallyMappableToU64,
     NumericalValue, StrColumn,
 };
 use common::{BitSet, TinySet};
@@ -26,7 +27,7 @@ use crate::aggregation::intermediate_agg_result::{
     IntermediateKey, IntermediateTermBucketEntry, IntermediateTermBucketResult,
 };
 use crate::aggregation::segment_agg_result::{BucketIdProvider, SegmentAggregationCollector};
-use crate::aggregation::{format_date, BucketId, Key};
+use crate::aggregation::{format_date, BucketId, Key, ValueSource};
 use crate::error::DataCorruption;
 use crate::TantivyError;
 
@@ -35,25 +36,23 @@ mod flattened_term_histogram;
 /// Contains all information required by the SegmentTermCollector to perform the
 /// terms aggregation on a segment.
 #[derive(Debug, Clone)]
-pub struct TermsAggReqData {
+pub(crate) struct TermsAggReqData {
     /// The column accessor to access the fast field values.
-    pub accessor: Column<u64>,
-    /// The type of the column.
-    pub column_type: ColumnType,
+    pub(crate) accessor: Arc<dyn ValueSource>,
     /// The string dictionary column if the field is of type text.
-    pub str_dict_column: Option<StrColumn>,
+    pub(crate) str_dict_column: Option<StrColumn>,
     /// The missing value as u64 value.
-    pub missing_value_for_accessor: Option<u64>,
+    pub(crate) missing_value_for_accessor: Option<u64>,
     /// Used to build the correct nested result when we have an empty result.
-    pub sug_aggregations: Aggregations,
+    pub(crate) sub_aggregations: Aggregations,
     /// The name of the aggregation.
-    pub name: String,
+    pub(crate) name: String,
     /// The normalized term aggregation request.
-    pub req: TermsAggregationInternal,
+    pub(crate) req: TermsAggregationInternal,
     /// Preloaded allowed term ords (string columns only). If set, only ords present are collected.
-    pub allowed_term_ids: Option<BitSet>,
+    pub(crate) allowed_term_ids: Option<BitSet>,
     /// True if this terms aggregation is at the top level of the aggregation tree (not nested).
-    pub is_top_level: bool,
+    pub(crate) is_top_level: bool,
 }
 
 impl TermsAggReqData {
@@ -394,7 +393,7 @@ pub(crate) fn build_segment_term_collector(
     node: &AggRefNode,
 ) -> crate::Result<Box<dyn SegmentAggregationCollector>> {
     let terms_req_data = req_data.get_term_req_data(node.idx_in_req_data).clone();
-    let column_type = terms_req_data.column_type;
+    let column_type = terms_req_data.accessor.column_type();
 
     if column_type == ColumnType::Bytes {
         return Err(TantivyError::InvalidArgument(format!(
@@ -423,7 +422,11 @@ pub(crate) fn build_segment_term_collector(
 
     // Let's see if we can use a vec to aggregate our data
     // instead of a hashmap.
-    let col_max_value = terms_req_data.accessor.max_value();
+    let col_max_value = terms_req_data
+        .accessor
+        .as_column()
+        .map(|col| col.max_value())
+        .unwrap_or(u64::MAX);
     let max_column_val: u64 =
         col_max_value.max(terms_req_data.missing_value_for_accessor.unwrap_or(0u64));
 
@@ -1064,7 +1067,7 @@ impl<TermMap: TermAggregationMap, B: SubAggBuffer> SegmentAggregationCollector
             .column_block_accessor
             .fetch_block_with_missing_unique_per_doc(
                 docs,
-                &req_data.accessor,
+                &*req_data.accessor,
                 req_data.missing_value_for_accessor,
                 false,
             );
@@ -1331,7 +1334,8 @@ where
         let mut out: Vec<(IntermediateKey, IntermediateTermBucketEntry)> =
             Vec::with_capacity(entries.len());
 
-        if term_req.column_type == ColumnType::Str {
+        let column_type = term_req.accessor.column_type();
+        if column_type == ColumnType::Str {
             let fallback_dict = Dictionary::empty();
             let term_dict = term_req
                 .str_dict_column
@@ -1390,7 +1394,7 @@ where
                 // TODO: Handle rev streaming for descending sorting by keys
                 let mut stream = term_dict.stream()?;
                 let empty_sub_aggregation =
-                    IntermediateAggregationResults::empty_from_req(&term_req.sug_aggregations);
+                    IntermediateAggregationResults::empty_from_req(&term_req.sub_aggregations);
                 while stream.advance() {
                     if dict.len() >= term_req.req.segment_size as usize {
                         break;
@@ -1418,7 +1422,7 @@ where
             }
 
             out.extend(dict);
-        } else if term_req.column_type == ColumnType::DateTime {
+        } else if column_type == ColumnType::DateTime {
             for (val, doc_count) in entries {
                 let intermediate_entry = into_intermediate_bucket_entry(
                     doc_count,
@@ -1429,7 +1433,7 @@ where
                 let date = format_date(val)?;
                 out.push((IntermediateKey::Str(date), intermediate_entry));
             }
-        } else if term_req.column_type == ColumnType::Bool {
+        } else if column_type == ColumnType::Bool {
             for (val, doc_count) in entries {
                 let intermediate_entry = into_intermediate_bucket_entry(
                     doc_count,
@@ -1439,9 +1443,17 @@ where
                 let val = bool::from_u64(val);
                 out.push((IntermediateKey::Bool(val), intermediate_entry));
             }
-        } else if term_req.column_type == ColumnType::IpAddr {
+        } else if column_type == ColumnType::IpAddr {
             let compact_space_accessor = term_req
                 .accessor
+                .as_column()
+                .ok_or_else(|| {
+                    TantivyError::AggregationError(
+                        crate::aggregation::AggregationError::InternalError(
+                            "IpAddr term keys require a physical column".to_string(),
+                        ),
+                    )
+                })?
                 .values
                 .clone()
                 .downcast_arc::<CompactSpaceU64Accessor>()
@@ -1464,7 +1476,7 @@ where
                 let val = Ipv6Addr::from_u128(val);
                 out.push((IntermediateKey::IpAddr(val), intermediate_entry));
             }
-        } else if term_req.column_type == ColumnType::F64 {
+        } else if column_type == ColumnType::F64 {
             // -0.0 and +0.0 both normalize to I64(0). Sort by normalized key to merge their
             // buckets below. Other distinct f64 encodings, including NaNs, remain distinct:
             // NaNs are not normalized and IntermediateKey compares them using total_cmp.
@@ -1497,13 +1509,13 @@ where
                     reborrow_opt_collector(&mut sub_agg_collector),
                     agg_data,
                 )?;
-                let key_val: NumericalValue = match term_req.column_type {
+                let key_val: NumericalValue = match column_type {
                     ColumnType::U64 => val.into(),
                     ColumnType::I64 => i64::from_u64(val).into(),
                     _ => {
                         return Err(TantivyError::SchemaError(format!(
                             "unknown key type: {}",
-                            term_req.column_type
+                            column_type
                         )))
                     }
                 };
@@ -1554,7 +1566,7 @@ pub(crate) trait GetDocCount {
     fn doc_count(&self) -> u64;
 }
 
-impl GetDocCount for (String, IntermediateTermBucketEntry) {
+impl<K> GetDocCount for (K, IntermediateTermBucketEntry) {
     fn doc_count(&self) -> u64 {
         self.1.doc_count
     }
diff --git a/src/aggregation/bucket/term_missing_agg.rs b/src/aggregation/bucket/term_missing_agg.rs
index 72dacfd07..1da26d856 100644
--- a/src/aggregation/bucket/term_missing_agg.rs
+++ b/src/aggregation/bucket/term_missing_agg.rs
@@ -20,13 +20,13 @@ use crate::aggregation::BucketId;
 /// - The field is not text and missing is provided as string (we cannot use the numeric missing
 ///   value optimization)
 #[derive(Default)]
-pub struct MissingTermAggReqData {
+pub(crate) struct MissingTermAggReqData {
     /// The accessors to check for existence of a value.
-    pub accessors: Vec<(Column<u64>, ColumnType)>,
+    pub(crate) accessors: Vec<(Column<u64>, ColumnType)>,
     /// The name of the aggregation.
-    pub name: String,
+    pub(crate) name: String,
     /// The original terms aggregation request.
-    pub req: TermsAggregation,
+    pub(crate) req: TermsAggregation,
 }
 
 impl MissingTermAggReqData {
diff --git a/src/aggregation/intermediate_agg_result.rs b/src/aggregation/intermediate_agg_result.rs
index ceb149b1c..fd9877e33 100644
--- a/src/aggregation/intermediate_agg_result.rs
+++ b/src/aggregation/intermediate_agg_result.rs
@@ -1213,19 +1213,20 @@ trait MergeFruits {
     fn merge_fruits(&mut self, other: Self) -> crate::Result<()>;
 }
 
-fn merge_maps<V: MergeFruits + Clone, T: Eq + PartialEq + Hash>(
+fn merge_maps<V: MergeFruits, T: Eq + Hash>(
     entries_left: &mut FxHashMap<T, V>,
-    mut entries_right: FxHashMap<T, V>,
+    entries_right: FxHashMap<T, V>,
 ) -> crate::Result<()> {
-    for (name, entry_left) in entries_left.iter_mut() {
-        if let Some(entry_right) = entries_right.remove(name) {
-            entry_left.merge_fruits(entry_right)?;
+    // Visit incoming entries, not the growing accumulator, so folding many results
+    // does not repeatedly hash all previously merged keys.
+    for (key, entry_right) in entries_right {
+        match entries_left.entry(key) {
+            Entry::Occupied(mut entry) => entry.get_mut().merge_fruits(entry_right)?,
+            Entry::Vacant(entry) => {
+                entry.insert(entry_right);
+            }
         }
     }
-
-    for (key, res) in entries_right.into_iter() {
-        entries_left.entry(key).or_insert(res);
-    }
     Ok(())
 }
 
@@ -1698,6 +1699,106 @@ mod tests {
         assert_range_trees_eq(&tree_left, &tree_expected);
     }
 
+    #[test]
+    fn test_multi_terms_repeated_merge_and_prune() {
+        let key = |id: u64| {
+            vec![
+                IntermediateKey::Str(format!("host-{}", id % 2)),
+                IntermediateKey::U64(id / 2),
+            ]
+        };
+        let make_result =
+            |data: &[(u64, u64)], source: &str| IntermediateBucketResult::MultiTerms {
+                buckets: IntermediateMultiTermsBucketResult {
+                    entries: data
+                        .iter()
+                        .map(|&(id, count)| {
+                            (
+                                key(id),
+                                IntermediateTermBucketEntry {
+                                    doc_count: count,
+                                    sub_aggregation: get_sub_test_tree(&[
+                                        ("shared".to_string(), count),
+                                        (source.to_string(), count),
+                                    ]),
+                                },
+                            )
+                        })
+                        .collect(),
+                    sum_other_doc_count: 2,
+                    doc_count_error_upper_bound: 1,
+                },
+            };
+        let mut merged = IntermediateBucketResult::MultiTerms {
+            buckets: Default::default(),
+        };
+        merged
+            .merge_fruits(make_result(&[(0, 3), (1, 5)], "first"))
+            .unwrap();
+        merged
+            .merge_fruits(make_result(&[(1, 7), (2, 11)], "second"))
+            .unwrap();
+        // A distributed fold may resume after serializing an intermediate result.
+        merged = postcard::from_bytes(&postcard::to_allocvec(&merged).unwrap()).unwrap();
+        merged.merge_fruits(make_result(&[], "empty")).unwrap();
+        merged
+            .merge_fruits(make_result(&[(0, 13), (3, 17)], "third"))
+            .unwrap();
+
+        let IntermediateBucketResult::MultiTerms { buckets } = &mut merged else {
+            panic!("expected multi_terms");
+        };
+        assert_eq!(buckets.entries.len(), 4);
+        assert_eq!(buckets.sum_other_doc_count, 8);
+        assert_eq!(buckets.doc_count_error_upper_bound, 4);
+        for (id, count, sources) in [
+            (0, 16, vec![("first", 3), ("third", 13)]),
+            (1, 12, vec![("first", 5), ("second", 7)]),
+            (2, 11, vec![("second", 11)]),
+            (3, 17, vec![("third", 17)]),
+        ] {
+            let entry = &buckets.entries[&key(id)];
+            assert_eq!(entry.doc_count, count);
+            let mut expected = vec![("shared".to_string(), count)];
+            expected.extend(
+                sources
+                    .into_iter()
+                    .map(|(name, count)| (name.to_string(), count)),
+            );
+            assert_range_trees_eq(&entry.sub_aggregation, &get_sub_test_tree(&expected));
+        }
+
+        let req = serde_json::from_value(serde_json::json!({
+            "terms": [{"field": "host"}, {"field": "path"}],
+            "size": 1,
+            "segment_size": 2
+        }))
+        .unwrap();
+        buckets
+            .prune_intermediate_results(&req, &Default::default(), PruneMode::Intermediate)
+            .unwrap();
+        assert_eq!(buckets.entries.len(), 2);
+        assert_eq!(buckets.sum_other_doc_count, 31); // 8 + 12 + 11
+        assert_eq!(buckets.doc_count_error_upper_bound, 16); // 4 + cutoff 12
+
+        // A previously pruned key can be inserted again in a subsequent fold.
+        merged
+            .merge_fruits(make_result(&[(2, 19)], "fourth"))
+            .unwrap();
+        let IntermediateBucketResult::MultiTerms { buckets } = &mut merged else {
+            unreachable!();
+        };
+        assert_eq!(buckets.entries.len(), 3);
+        assert_eq!(buckets.entries[&key(2)].doc_count, 19);
+        buckets
+            .prune_intermediate_results(&req, &Default::default(), PruneMode::Final)
+            .unwrap();
+        assert_eq!(buckets.entries.len(), 1);
+        assert_eq!(buckets.entries[&key(2)].doc_count, 19);
+        assert_eq!(buckets.sum_other_doc_count, 66); // 31 + 2 + 16 + 17
+        assert_eq!(buckets.doc_count_error_upper_bound, 17); // 16 + 1, no final cutoff
+    }
+
     #[test]
     fn test_prune_intermediate_results_finalizer_size() {
         use crate::aggregation::bucket::TermsAggregation;
diff --git a/src/aggregation/metric/cardinality.rs b/src/aggregation/metric/cardinality.rs
deleted file mode 100644
index ca0a00fb2..000000000
--- a/src/aggregation/metric/cardinality.rs
+++ /dev/null
@@ -1,1407 +0,0 @@
-use std::fmt::Debug;
-use std::hash::Hash;
-use std::io;
-
-use columnar::column_values::CompactSpaceU64Accessor;
-use columnar::{Column, ColumnType, Dictionary, StrColumn};
-use common::{BitSet, TinySet};
-use datasketches::hll::{Coupon, HllSketch, HllType, HllUnion};
-use rustc_hash::{FxBuildHasher, FxHashMap, FxHashSet};
-use serde::{Deserialize, Deserializer, Serialize, Serializer};
-
-use crate::aggregation::agg_data::AggregationsSegmentCtx;
-use crate::aggregation::intermediate_agg_result::{
-    IntermediateAggregationResult, IntermediateAggregationResults, IntermediateMetricResult,
-};
-use crate::aggregation::segment_agg_result::SegmentAggregationCollector;
-use crate::aggregation::*;
-use crate::TantivyError;
-
-/// Log2 of the number of registers for the HLL sketch.
-/// 2^11 = 2048 registers, giving ~2.3% relative error and ~1KB per sketch (Hll4).
-const LG_K: u8 = 11;
-
-/// Promote FxHashSet<u64> -> PagedBitset at ~3% density (`len * 32 >
-/// dict_num_terms`). Past this point the bitset (~`dict_num_terms / 7.5`
-/// bytes) is smaller than the hashset (~10 B/entry minimum) and avoids
-/// the per-insert hash.
-const PROMOTION_RATIO: u64 = 32;
-
-/// # Cardinality
-///
-/// The cardinality aggregation allows for computing an estimate
-/// of the number of different values in a data set based on the
-/// Apache DataSketches HyperLogLog algorithm. This is particularly useful for
-/// understanding the uniqueness of values in a large dataset where counting
-/// each unique value individually would be computationally expensive.
-///
-/// For example, you might use a cardinality aggregation to estimate the number
-/// of unique visitors to a website by aggregating on a field that contains
-/// user IDs or session IDs.
-///
-/// To use the cardinality aggregation, you'll need to provide a field to
-/// aggregate on. The following example demonstrates a request for the cardinality
-/// of the "user_id" field:
-///
-/// ```JSON
-/// {
-///     "cardinality": {
-///         "field": "user_id"
-///     }
-/// }
-/// ```
-///
-/// This request will return an estimate of the number of unique values in the
-/// "user_id" field.
-///
-/// ## Missing Values
-///
-/// The `missing` parameter defines how documents that are missing a value should be treated.
-/// By default, documents without a value for the specified field are ignored. However, you can
-/// specify a default value for these documents using the `missing` parameter. This can be useful
-/// when you want to include documents with missing values in the aggregation.
-///
-/// For example, the following request treats documents with missing values in the "user_id"
-/// field as if they had a value of "unknown":
-///
-/// ```JSON
-/// {
-///     "cardinality": {
-///         "field": "user_id",
-///         "missing": "unknown"
-///     }
-/// }
-/// ```
-///
-/// # Estimation Accuracy
-///
-/// The cardinality aggregation provides an approximate count, which is usually
-/// accurate within a small error range. This trade-off allows for efficient
-/// computation even on very large datasets.
-#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
-pub struct CardinalityAggregationReq {
-    /// The field name to compute the percentiles on.
-    pub field: String,
-    /// The missing parameter defines how documents that are missing a value should be treated.
-    /// By default they will be ignored but it is also possible to treat them as if they had a
-    /// value. Examples in JSON format:
-    /// { "field": "my_numbers", "missing": "10.0" }
-    #[serde(skip_serializing_if = "Option::is_none", default)]
-    pub missing: Option<Key>,
-}
-
-/// Contains all information required by the SegmentCardinalityCollector to perform the
-/// cardinality aggregation on a segment.
-pub struct CardinalityAggReqData {
-    /// The column accessor to access the fast field values.
-    pub accessor: Column<u64>,
-    /// The column_type of the field.
-    pub column_type: ColumnType,
-    /// The string dictionary column if the field is of type string.
-    pub str_dict_column: Option<StrColumn>,
-    /// The missing value normalized to the internal u64 representation of the field type.
-    pub missing_value_for_accessor: Option<u64>,
-    /// The name of the aggregation.
-    pub name: String,
-    /// The aggregation request.
-    pub req: CardinalityAggregationReq,
-}
-
-impl CardinalityAggReqData {
-    /// Estimate the memory consumption of this struct in bytes.
-    pub fn get_memory_consumption(&self) -> usize {
-        std::mem::size_of::<Self>()
-    }
-}
-
-impl CardinalityAggregationReq {
-    /// Creates a new [`CardinalityAggregationReq`] instance from a field name.
-    pub fn from_field_name(field_name: String) -> Self {
-        Self {
-            field: field_name,
-            missing: None,
-        }
-    }
-    /// Returns the field name the aggregation is computed on.
-    pub fn field_name(&self) -> &str {
-        &self.field
-    }
-}
-
-/// A CouponCache is here to cache the mapping term ordinal -> coupon (see above).
-/// The idea is that we do not want to fetch terms associated to several term ordinals,
-/// several times due to the fact that we have several buckets.
-enum CouponCache {
-    Dense {
-        coupon_map: Vec<Coupon>,
-        missing_coupon_opt: Option<Coupon>,
-    },
-    Sparse {
-        coupon_map: FxHashMap<u64, Coupon>,
-        missing_coupon_opt: Option<Coupon>,
-    },
-}
-
-impl CouponCache {
-    fn new(
-        term_ords: Vec<u64>,
-        coupons: Vec<Coupon>,
-        missing_coupon_opt: Option<Coupon>,
-    ) -> CouponCache {
-        let num_terms = term_ords.len();
-        assert_eq!(num_terms, coupons.len());
-        if term_ords.is_empty() {
-            return CouponCache::Dense {
-                coupon_map: Vec::new(),
-                missing_coupon_opt,
-            };
-        }
-        let highest_term_ord = term_ords.last().copied().unwrap_or(0u64);
-        // We prefer the dense implementation, if it is not too wasteful.
-        // There are two cases for which we can use it.
-        // 1- if the data is small.
-        // 2- if the data is not necessarily small, but due to a high occupancy ratio, the RAM usage
-        // is not that much bigger than if we had used a HashSet. (occupancy ratio + extra
-        // metadata ~ x2.25)
-        let should_use_dense =
-            highest_term_ord < 1_000_000u64 || highest_term_ord < num_terms as u64 * 3u64;
-        if should_use_dense {
-            // We don't really care about the value here. We will populate all the values we will
-            // read anyway.
-            let uninitialized_coupon = Coupon::from_hash(0);
-            let mut coupon_map: Vec<Coupon> =
-                vec![uninitialized_coupon; highest_term_ord as usize + 1];
-
-            for (term_ord, coupon) in term_ords.into_iter().zip(coupons) {
-                coupon_map[term_ord as usize] = coupon;
-            }
-            CouponCache::Dense {
-                coupon_map,
-                missing_coupon_opt,
-            }
-        } else {
-            let coupon_map: FxHashMap<u64, Coupon> = term_ords.into_iter().zip(coupons).collect();
-            CouponCache::Sparse {
-                coupon_map,
-                missing_coupon_opt,
-            }
-        }
-    }
-}
-
-// =================================================================
-// PagedBitset: a sparse bitset indexed by term_ord.
-//
-// Used as the dense alternative to FxHashSet<u64> once a string
-// cardinality bucket has accumulated enough unique term ordinals.
-// Memory is bounded to (touched pages) * (page bytes), not
-// (max_term_ord / 8).
-//
-// Page geometry mirrors `PagedTermMap` in `term_agg.rs`: 1024 ords
-// per page, lazy `Vec<Option<Box<Page>>>` directory.
-// =================================================================
-const BITSET_PAGE_SHIFT: u32 = 10;
-const BITSET_PAGE_BITS: u64 = 1u64 << BITSET_PAGE_SHIFT; // 1024
-const BITSET_PAGE_MASK: u64 = BITSET_PAGE_BITS - 1;
-const BITSET_WORDS_PER_PAGE: usize = (BITSET_PAGE_BITS / 64) as usize; // 16
-
-#[derive(Clone)]
-struct PagedBitsetPage {
-    words: [TinySet; BITSET_WORDS_PER_PAGE],
-}
-
-impl PagedBitsetPage {
-    fn new() -> Self {
-        Self {
-            words: [TinySet::empty(); BITSET_WORDS_PER_PAGE],
-        }
-    }
-}
-
-pub(crate) struct PagedBitset {
-    pages: Vec<Option<Box<PagedBitsetPage>>>,
-    /// Cached number of set bits, maintained on insert.
-    count: u64,
-}
-
-impl PagedBitset {
-    /// Allocates a directory big enough to hold ords up to and including
-    /// `max_term_ord`. Pages are allocated lazily on first set.
-    fn with_max_term_ord(max_term_ord: u64) -> Self {
-        let max_page_idx = (max_term_ord >> BITSET_PAGE_SHIFT) as usize;
-        let num_pages = max_page_idx + 1;
-        Self {
-            pages: vec![None; num_pages],
-            count: 0,
-        }
-    }
-
-    #[inline]
-    fn insert(&mut self, term_ord: u64) {
-        let page_idx = (term_ord >> BITSET_PAGE_SHIFT) as usize;
-        let intra = term_ord & BITSET_PAGE_MASK;
-        let word_idx = (intra >> 6) as usize;
-        let bit_idx = (intra & 63) as u32;
-
-        let page = match &mut self.pages[page_idx] {
-            Some(p) => p,
-            None => {
-                self.pages[page_idx] = Some(Box::new(PagedBitsetPage::new()));
-                self.pages[page_idx].as_mut().unwrap()
-            }
-        };
-        if page.words[word_idx].insert_mut(bit_idx) {
-            self.count += 1;
-        }
-    }
-
-    /// Number of set bits. O(1).
-    #[inline]
-    fn len(&self) -> u64 {
-        self.count
-    }
-
-    /// Iterate set ords in ascending order.
-    fn iter_sorted(&self) -> impl Iterator<Item = u64> + '_ {
-        self.pages
-            .iter()
-            .enumerate()
-            .filter_map(|(page_idx, page_opt)| page_opt.as_ref().map(|p| (page_idx, p)))
-            .flat_map(|(page_idx, page)| {
-                let page_base_ord = (page_idx as u64) << BITSET_PAGE_SHIFT;
-                page.words
-                    .iter()
-                    .enumerate()
-                    .flat_map(move |(word_idx, &word)| {
-                        let word_base_ord = page_base_ord + (word_idx as u64) * 64;
-                        word.into_iter()
-                            .map(move |bit| word_base_ord + u64::from(bit))
-                    })
-            })
-    }
-}
-
-/// Threshold below which we use `BitSet` instead of `TermOrdSet`.
-///
-/// Both `BitSet` and `FxHashSet<u64>` have the same 32-byte struct, so the comparison is heap only:
-///   * `BitSet` at T=256: 5 `TinySet` words covering 258 bits (with the missing-value sentinel) =
-///     40 bytes.
-///   * `FxHashSet<u64>` after one insert: 4-bucket hashbrown table ≈ 56 bytes
-pub(crate) const BITSET_MAX_TERM_ORD: u64 = 256;
-
-// =================================================================
-// TermOrdAccumulator: per-bucket abstraction over the entries set.
-//
-// Implementations:
-//   - `BitSet` (from `common`): used when `column.max_value()` is small (< BITSET_MAX_TERM_ORD).
-//     Pre-allocated, no promotion.
-//   - `TermOrdSet`: adaptive, starts as FxHashSet and promotes to a paged bitset when occupancy
-//     crosses the density threshold (only if promotion is enabled — typically gated on top-level
-//     aggregation).
-//
-// The trait lets `SegmentCardinalityCollector` be generic over the choice
-// so the hot collect() loop monomorphizes to a direct call (no enum
-// dispatch per insert).
-// =================================================================
-pub(crate) trait TermOrdAccumulator: Sized {
-    /// Construct an empty accumulator.
-    /// `max_term_ord_inclusive` is the largest term_ord that may be
-    /// inserted (used to size pre-allocated bitsets and the dense bitset
-    /// on promotion).
-    fn new(max_term_ord_inclusive: u64) -> Self;
-    fn insert(&mut self, term_ord: u64);
-    /// Bulk insert. Implementations may override to hoist any inner
-    /// dispatch outside the loop. Default loops `insert`.
-    #[inline]
-    fn extend_from_iter<I: IntoIterator<Item = u64>>(&mut self, ords: I) {
-        for ord in ords {
-            self.insert(ord);
-        }
-    }
-    /// Hook called once per ingested block. Adaptive impls use this to
-    /// decide on sparse->dense promotion.
-    fn maybe_compact(&mut self) {}
-    fn len(&self) -> usize;
-    fn iter_ords(&self) -> impl Iterator<Item = u64> + '_;
-}
-
-impl TermOrdAccumulator for BitSet {
-    #[inline]
-    fn new(max_term_ord_inclusive: u64) -> Self {
-        // `BitSet::with_max_value(M)` accepts ords in [0, M).
-        // We need ords up to and including `max_term_ord_inclusive`, plus
-        // the missing-value sentinel `column.max_value() + 1`.
-        BitSet::with_max_value((max_term_ord_inclusive + 2) as u32)
-    }
-    #[inline]
-    fn insert(&mut self, term_ord: u64) {
-        BitSet::insert(self, term_ord as u32);
-    }
-    #[inline]
-    fn len(&self) -> usize {
-        BitSet::len(self)
-    }
-    fn iter_ords(&self) -> impl Iterator<Item = u64> + '_ {
-        // `BitSet` itself doesn't expose iteration, but
-        // `BitSet::tinyset(bucket)` does. Walk per-bucket and yield each
-        // set bit. The capacity is `max_value()`; iterating to
-        // `div_ceil(64)` covers every possible ord exactly once.
-        let num_buckets = self.max_value().div_ceil(64);
-        (0..num_buckets).flat_map(move |bucket| {
-            let chunk_base = u64::from(bucket) * 64;
-            self.tinyset(bucket)
-                .into_iter()
-                .map(move |bit| chunk_base + u64::from(bit))
-        })
-    }
-}
-
-// =================================================================
-// TermOrdSet: adaptive sparse->dense accumulator.
-//
-// Starts as an FxHashSet (cheap when few ords are seen). When occupancy
-// crosses `len * PROMOTION_RATIO > max_term_ord_inclusive`, drains into
-// a `PagedBitset` and continues dense. Promotion is one-way.
-// =================================================================
-pub(crate) struct TermOrdSet {
-    inner: TermOrdSetInner,
-    /// Largest term_ord that may be inserted. Used for both sizing the
-    /// dense bitset on promotion and as the promotion-threshold reference.
-    max_term_ord_inclusive: u64,
-}
-
-enum TermOrdSetInner {
-    Sparse(FxHashSet<u64>),
-    Dense(PagedBitset),
-}
-
-impl TermOrdAccumulator for TermOrdSet {
-    fn new(max_term_ord_inclusive: u64) -> Self {
-        Self {
-            inner: TermOrdSetInner::Sparse(FxHashSet::default()),
-            max_term_ord_inclusive,
-        }
-    }
-
-    #[inline]
-    fn insert(&mut self, term_ord: u64) {
-        match &mut self.inner {
-            TermOrdSetInner::Sparse(set) => {
-                set.insert(term_ord);
-            }
-            TermOrdSetInner::Dense(bitset) => bitset.insert(term_ord),
-        }
-    }
-
-    /// Hoist the Sparse/Dense match outside the per-ord loop so that a
-    /// block of inserts dispatches once.
-    fn extend_from_iter<I: IntoIterator<Item = u64>>(&mut self, ords: I) {
-        match &mut self.inner {
-            TermOrdSetInner::Sparse(set) => {
-                for ord in ords {
-                    set.insert(ord);
-                }
-            }
-            TermOrdSetInner::Dense(bitset) => {
-                for ord in ords {
-                    bitset.insert(ord);
-                }
-            }
-        }
-    }
-
-    fn maybe_compact(&mut self) {
-        let TermOrdSetInner::Sparse(set) = &mut self.inner else {
-            return;
-        };
-        if set.len() as u64 * PROMOTION_RATIO <= self.max_term_ord_inclusive {
-            return;
-        }
-        // Size for ord <= max_term_ord_inclusive plus the missing sentinel
-        // (column.max_value() + 1, which may equal max_term_ord_inclusive
-        // when the column references every dictionary term).
-        let mut bitset = PagedBitset::with_max_term_ord(self.max_term_ord_inclusive + 1);
-        let set = std::mem::take(set);
-        for ord in set {
-            bitset.insert(ord);
-        }
-        self.inner = TermOrdSetInner::Dense(bitset);
-    }
-
-    fn len(&self) -> usize {
-        match &self.inner {
-            TermOrdSetInner::Sparse(set) => set.len(),
-            TermOrdSetInner::Dense(bitset) => bitset.len() as usize,
-        }
-    }
-
-    fn iter_ords(&self) -> impl Iterator<Item = u64> + '_ {
-        match &self.inner {
-            TermOrdSetInner::Sparse(set) => itertools::Either::Left(set.iter().copied()),
-            TermOrdSetInner::Dense(bitset) => itertools::Either::Right(bitset.iter_sorted()),
-        }
-    }
-}
-
-pub(crate) struct SegmentCardinalityCollector<S: TermOrdAccumulator> {
-    /// Buckets are Some(_) until they get consumed by into_intermediate_results().
-    buckets: Vec<Option<SegmentCardinalityCollectorBucket<S>>>,
-    accessor_idx: usize,
-    /// The column accessor to access the fast field values.
-    accessor: Column<u64>,
-    /// The column_type of the field.
-    column_type: ColumnType,
-    /// The missing value normalized to the internal u64 representation of the field type.
-    missing_value_for_accessor: Option<u64>,
-    coupon_cache: Option<CouponCache>,
-    /// Largest term_ord that may be inserted into a bucket. For str columns
-    /// this is `accessor.max_value()`; for non-str columns this is unused
-    /// (no inserts go into `entries`) and set to 0.
-    max_term_ord_inclusive: u64,
-}
-
-impl<S: TermOrdAccumulator> Debug for SegmentCardinalityCollector<S> {
-    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
-        f.debug_struct("SegmentCardinalityCollector")
-            .field("column_type", &self.column_type)
-            .field(
-                "missing_value_for_accessor",
-                &self.missing_value_for_accessor,
-            )
-            .finish()
-    }
-}
-
-/// Per-bucket state. Shape depends on column kind: str columns dedup
-/// term ords and only build the HLL sketch at finalization (saves the
-/// ~96 B `CardinalityCollector` per bucket during collect); numeric/IpAddr
-/// columns feed the sketch directly during collect.
-pub(crate) enum SegmentCardinalityCollectorBucket<S: TermOrdAccumulator> {
-    Str(S),
-    Numeric(CardinalityCollector),
-}
-impl<S: TermOrdAccumulator> SegmentCardinalityCollectorBucket<S> {
-    #[inline(always)]
-    pub fn new(column_type: ColumnType, max_term_ord_inclusive: u64) -> Self {
-        if column_type == ColumnType::Str {
-            Self::Str(S::new(max_term_ord_inclusive))
-        } else {
-            Self::Numeric(CardinalityCollector::new(column_type as u8))
-        }
-    }
-
-    // Returns a intermediate metric result.
-    //
-    // If the column is not str, the values have been added to the
-    // sketch during collection.
-    //
-    // If the column is str, then the values are dictionary encoded
-    // and have not been added to the sketch yet.
-    // We need to resolves the term ords accumulated in the str entries
-    // with the coupon cache, and append the results to a fresh sketch.
-    fn into_intermediate_metric_result(
-        self,
-        coupon_cache_opt: Option<&CouponCache>,
-    ) -> crate::Result<IntermediateMetricResult> {
-        let cardinality = match self {
-            Self::Str(entries) => {
-                let mut cardinality = CardinalityCollector::new(ColumnType::Str as u8);
-                if let Some(coupon_cache) = coupon_cache_opt {
-                    // Sketch must be empty for str columns: coupons are appended here
-                    // from the term_ord set (and not directly during collection).
-                    assert!(cardinality.sketch.is_empty());
-                    append_to_sketch(&entries, coupon_cache, &mut cardinality);
-                }
-                cardinality
-            }
-            Self::Numeric(cardinality) => cardinality,
-        };
-        Ok(IntermediateMetricResult::Cardinality(cardinality))
-    }
-}
-
-/// Builds a coupon cache from the given buckets, dictionary, and optional missing value.
-/// Returns a mapping from term_ord to the hash (coupon) of the associated term.
-fn build_coupon_cache<S: TermOrdAccumulator>(
-    buckets: &[Option<SegmentCardinalityCollectorBucket<S>>],
-    dictionary: &Dictionary,
-    missing_value_opt: Option<&Key>,
-) -> io::Result<CouponCache> {
-    // Caller restricts this to str cardinality collectors, so every
-    // present bucket must be the `Str` variant. Pass 1 validates and
-    // computes the capacity hint; pass 2 inserts.
-    let mut max_bucket_len = 0usize;
-    for bucket in buckets.iter().flatten() {
-        match bucket {
-            SegmentCardinalityCollectorBucket::Str(entries) => {
-                max_bucket_len = max_bucket_len.max(entries.len());
-            }
-            SegmentCardinalityCollectorBucket::Numeric(_) => {
-                return Err(io::Error::other(
-                    "build_coupon_cache invoked with a non-str bucket",
-                ));
-            }
-        }
-    }
-    let mut term_ords_set = FxHashSet::with_capacity_and_hasher(max_bucket_len * 2, FxBuildHasher);
-    for bucket in buckets.iter().flatten() {
-        if let SegmentCardinalityCollectorBucket::Str(entries) = bucket {
-            term_ords_set.extend(entries.iter_ords());
-        }
-    }
-    let mut term_ords: Vec<u64> = term_ords_set.into_iter().collect();
-    term_ords.sort_unstable();
-
-    term_ords.pop_if(|highest_term_ord| *highest_term_ord >= dictionary.num_terms() as u64);
-
-    let mut coupons: Vec<Coupon> = Vec::with_capacity(term_ords.len());
-    let all_term_ords_found: bool =
-        dictionary.sorted_ords_to_term_cb(&term_ords, |term_bytes| {
-            let coupon: Coupon = Coupon::from_hash(term_bytes);
-            coupons.push(coupon);
-        })?;
-    assert!(all_term_ords_found);
-
-    // Regardless of whether or not there is effectively a missing value in one of the buckets,
-    // we populate the cache with the missing key too (if any).
-    let missing_coupon_opt: Option<Coupon> = missing_value_opt.map(|missing_key| {
-        if let Key::Str(missing_value_str) = missing_key {
-            Coupon::from_hash(missing_value_str.as_bytes())
-        } else {
-            // See https://github.com/quickwit-oss/tantivy/issues/2891
-            // A missing key with a type different from Str will not work as intended
-            // for the moment.
-            //
-            // Right now this is just a partial workaround.
-            Coupon::from_hash("__tantivy_missing_non_str__".as_bytes())
-        }
-    });
-    Ok(CouponCache::new(term_ords, coupons, missing_coupon_opt))
-}
-
-fn append_to_sketch<S: TermOrdAccumulator>(
-    term_ords: &S,
-    coupon_cache: &CouponCache,
-    sketch: &mut CardinalityCollector,
-) {
-    match coupon_cache {
-        CouponCache::Dense {
-            coupon_map,
-            missing_coupon_opt,
-        } => {
-            for term_ord in term_ords.iter_ords() {
-                if let Some(coupon) = coupon_map
-                    .get(term_ord as usize)
-                    .copied()
-                    .or(*missing_coupon_opt)
-                {
-                    sketch.insert_coupon(coupon);
-                }
-            }
-        }
-        CouponCache::Sparse {
-            coupon_map,
-            missing_coupon_opt,
-        } => {
-            for term_ord in term_ords.iter_ords() {
-                if let Some(coupon) = coupon_map.get(&term_ord).copied().or(*missing_coupon_opt) {
-                    sketch.insert_coupon(coupon);
-                }
-            }
-        }
-    }
-}
-
-impl<S: TermOrdAccumulator> SegmentCardinalityCollector<S> {
-    pub fn from_req(
-        column_type: ColumnType,
-        accessor_idx: usize,
-        accessor: Column<u64>,
-        missing_value_for_accessor: Option<u64>,
-        max_term_ord_inclusive: u64,
-    ) -> Self {
-        Self {
-            buckets: Vec::new(),
-            column_type,
-            accessor_idx,
-            accessor,
-            missing_value_for_accessor,
-            coupon_cache: None,
-            max_term_ord_inclusive,
-        }
-    }
-
-    fn fetch_block_with_field(
-        &mut self,
-        docs: &[crate::DocId],
-        agg_data: &mut AggregationsSegmentCtx,
-    ) {
-        agg_data.column_block_accessor.fetch_block_with_missing(
-            docs,
-            &self.accessor,
-            self.missing_value_for_accessor,
-        );
-    }
-}
-
-impl<S: TermOrdAccumulator + 'static> SegmentAggregationCollector
-    for SegmentCardinalityCollector<S>
-{
-    fn add_intermediate_aggregation_result(
-        &mut self,
-        agg_data: &AggregationsSegmentCtx,
-        results: &mut IntermediateAggregationResults,
-        bucket_id: BucketId,
-    ) -> crate::Result<()> {
-        self.prepare_max_bucket(bucket_id, agg_data)?;
-        let req_data = &agg_data.get_cardinality_req_data(self.accessor_idx);
-        // Strings are dictionary encoded. Fetching the terms associated to strings
-        // is expensive. For this reason, we do that once for all buckets and cache the results
-        // here.
-        if let Some(str_dict_column) = &req_data.str_dict_column {
-            // Ensure the coupon cache is populated.
-            // A mapping from term_ord to the hash of the associated term.
-            // The missing value sentinel will be associated to the hash of the missing value if
-            // any.
-            if self.coupon_cache.is_none() {
-                self.coupon_cache = Some(build_coupon_cache(
-                    &self.buckets,
-                    str_dict_column.dictionary(),
-                    req_data.req.missing.as_ref(),
-                )?);
-            }
-        }
-        let name = req_data.name.to_string();
-        // take the bucket in buckets and replace it with a new empty one
-        let Some(bucket) = self.buckets[bucket_id as usize].take() else {
-            return Err(crate::TantivyError::InternalError(
-                "the same bucket should not be finalized twice.".to_string(),
-            ));
-        };
-        let intermediate_result =
-            bucket.into_intermediate_metric_result(self.coupon_cache.as_ref())?;
-        results.push(
-            name,
-            IntermediateAggregationResult::Metric(intermediate_result),
-        )?;
-
-        Ok(())
-    }
-
-    fn collect(
-        &mut self,
-        parent_bucket_id: BucketId,
-        docs: &[crate::DocId],
-        agg_data: &mut AggregationsSegmentCtx,
-    ) -> crate::Result<()> {
-        self.fetch_block_with_field(docs, agg_data);
-        let Some(bucket) = &mut self.buckets[parent_bucket_id as usize].as_mut() else {
-            return Err(crate::TantivyError::InternalError(
-                "collection should not happen after finalization".to_string(),
-            ));
-        };
-        let col_block_accessor = &agg_data.column_block_accessor;
-        match bucket {
-            SegmentCardinalityCollectorBucket::Str(entries) => {
-                // Promotion check runs on the pre-block state: the first call
-                // sees an empty set (no-op), and the last block of inserts
-                // doesn't trigger a promotion of a set we won't grow further.
-                // The trait dispatches once per block (via `extend_from_iter`)
-                // for adaptive variants and inlines to a tight loop for the
-                // BitSet path.
-                entries.maybe_compact();
-                entries.extend_from_iter(col_block_accessor.iter_vals());
-            }
-            SegmentCardinalityCollectorBucket::Numeric(cardinality) => {
-                if self.column_type == ColumnType::IpAddr {
-                    let compact_space_accessor = self
-                        .accessor
-                        .values
-                        .clone()
-                        .downcast_arc::<CompactSpaceU64Accessor>()
-                        .map_err(|_| {
-                            TantivyError::AggregationError(
-                                crate::aggregation::AggregationError::InternalError(
-                                    "Type mismatch: Could not downcast to CompactSpaceU64Accessor"
-                                        .to_string(),
-                                ),
-                            )
-                        })?;
-                    for val in col_block_accessor.iter_vals() {
-                        let val: u128 = compact_space_accessor.compact_to_u128(val as u32);
-                        cardinality.insert(val);
-                    }
-                } else {
-                    for val in col_block_accessor.iter_vals() {
-                        cardinality.insert(val);
-                    }
-                }
-            }
-        }
-
-        Ok(())
-    }
-
-    fn prepare_max_bucket(
-        &mut self,
-        max_bucket: BucketId,
-        _agg_data: &AggregationsSegmentCtx,
-    ) -> crate::Result<()> {
-        if max_bucket as usize >= self.buckets.len() {
-            let column_type = self.column_type;
-            let max_term_ord_inclusive = self.max_term_ord_inclusive;
-            self.buckets.resize_with(max_bucket as usize + 1, || {
-                Some(SegmentCardinalityCollectorBucket::<S>::new(
-                    column_type,
-                    max_term_ord_inclusive,
-                ))
-            });
-        }
-        Ok(())
-    }
-
-    fn compute_metric_value(
-        &self,
-        bucket_id: BucketId,
-        sub_agg_name: &str,
-        sub_agg_property: &str,
-        agg_data: &AggregationsSegmentCtx,
-    ) -> Option<f64> {
-        let req_data = &agg_data.get_cardinality_req_data(self.accessor_idx);
-        if req_data.name != sub_agg_name || !sub_agg_property.is_empty() {
-            return None;
-        }
-        let bucket = self.buckets.get(bucket_id as usize)?.as_ref()?;
-        // For string columns the sketch isn't built until finalization; the
-        // term_ord set's len is the exact distinct count. For numeric columns
-        // the sketch is populated during collect.
-        match bucket {
-            SegmentCardinalityCollectorBucket::Str(entries) => Some(entries.len() as f64),
-            SegmentCardinalityCollectorBucket::Numeric(cardinality) => {
-                Some(cardinality.sketch.estimate().trunc())
-            }
-        }
-    }
-}
-
-#[derive(Clone, Debug)]
-/// The cardinality collector used during segment collection and for merging results.
-/// Uses Apache DataSketches HLL (lg_k=11, Hll4) for compact binary serialization
-/// and cross-language compatibility (e.g. Java `datasketches` library).
-pub struct CardinalityCollector {
-    sketch: HllSketch,
-    /// Salt derived from `ColumnType`, used to differentiate values of different column types
-    /// that map to the same u64 (e.g. bool `false` = 0 vs i64 `0`).
-    /// Not serialized — only needed during insertion, not after sketch registers are populated.
-    salt: u8,
-}
-
-impl Default for CardinalityCollector {
-    fn default() -> Self {
-        Self::new(0)
-    }
-}
-
-impl PartialEq for CardinalityCollector {
-    fn eq(&self, _other: &Self) -> bool {
-        false
-    }
-}
-
-impl Serialize for CardinalityCollector {
-    fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
-        let bytes = self.sketch.serialize();
-        serializer.serialize_bytes(&bytes)
-    }
-}
-
-impl<'de> Deserialize<'de> for CardinalityCollector {
-    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
-        let bytes: Vec<u8> = Deserialize::deserialize(deserializer)?;
-        let sketch = HllSketch::deserialize(&bytes).map_err(serde::de::Error::custom)?;
-        Ok(Self { sketch, salt: 0 })
-    }
-}
-
-impl CardinalityCollector {
-    fn new(salt: u8) -> Self {
-        Self {
-            sketch: HllSketch::new(LG_K, HllType::Hll8),
-            salt,
-        }
-    }
-
-    /// Insert a value into the HLL sketch, salted by the column type.
-    /// The salt ensures that identical u64 values from different column types
-    /// (e.g. bool `false` vs i64 `0`) are counted as distinct.
-    fn insert<T: Hash>(&mut self, value: T) {
-        self.sketch.update((self.salt, value));
-    }
-
-    fn insert_coupon(&mut self, coupon: Coupon) {
-        self.sketch.update_with_coupon(coupon);
-    }
-
-    /// Compute the final cardinality estimate.
-    pub fn finalize(self) -> Option<f64> {
-        Some(self.sketch.estimate().trunc())
-    }
-
-    /// Serialize the HLL sketch to its compact binary representation.
-    /// The format is cross-language compatible with Apache DataSketches (Java, C++, Python).
-    pub fn to_sketch_bytes(&self) -> Vec<u8> {
-        self.sketch.serialize()
-    }
-
-    pub(crate) fn merge_fruits(&mut self, right: CardinalityCollector) -> crate::Result<()> {
-        let mut union = HllUnion::new(LG_K);
-        union.update(&self.sketch);
-        union.update(&right.sketch);
-        self.sketch = union.to_sketch(HllType::Hll8);
-        Ok(())
-    }
-}
-
-#[cfg(test)]
-mod tests {
-
-    use std::net::IpAddr;
-    use std::str::FromStr;
-
-    use columnar::MonotonicallyMappableToU64;
-
-    use crate::aggregation::agg_req::Aggregations;
-    use crate::aggregation::tests::{exec_request, get_test_index_from_terms};
-    use crate::schema::{IntoIpv6Addr, Schema, FAST, STRING};
-    use crate::Index;
-
-    #[test]
-    fn cardinality_aggregation_test_empty_index() -> crate::Result<()> {
-        let values = vec![];
-        let index = get_test_index_from_terms(false, &values)?;
-        let agg_req: Aggregations = serde_json::from_value(json!({
-            "cardinality": {
-                "cardinality": {
-                    "field": "string_id",
-                }
-            },
-        }))
-        .unwrap();
-
-        let res = exec_request(agg_req, &index)?;
-        assert_eq!(res["cardinality"]["value"], 0.0);
-
-        Ok(())
-    }
-
-    #[test]
-    fn cardinality_aggregation_test_single_segment() -> crate::Result<()> {
-        cardinality_aggregation_test_merge_segment(true)
-    }
-    #[test]
-    fn cardinality_aggregation_test() -> crate::Result<()> {
-        cardinality_aggregation_test_merge_segment(false)
-    }
-    fn cardinality_aggregation_test_merge_segment(merge_segments: bool) -> crate::Result<()> {
-        let segment_and_terms = vec![
-            vec!["terma"],
-            vec!["termb"],
-            vec!["termc"],
-            vec!["terma"],
-            vec!["terma"],
-            vec!["terma"],
-            vec!["termb"],
-            vec!["terma"],
-        ];
-        let index = get_test_index_from_terms(merge_segments, &segment_and_terms)?;
-        let agg_req: Aggregations = serde_json::from_value(json!({
-            "cardinality": {
-                "cardinality": {
-                    "field": "string_id",
-                }
-            },
-        }))
-        .unwrap();
-
-        let res = exec_request(agg_req, &index)?;
-        assert_eq!(res["cardinality"]["value"], 3.0);
-
-        Ok(())
-    }
-
-    /// Build a single-segment string-cardinality index with 32 unique terms.
-    /// `column.max_value() = 31` is well below `BITSET_MAX_TERM_ORD`,
-    /// so the bucket exercises the `BitSet` path end to end.
-    #[test]
-    fn cardinality_aggregation_test_str_bitset() -> crate::Result<()> {
-        let terms: Vec<String> = (0..32).map(|i| format!("term_{i}")).collect();
-        let term_refs: Vec<Vec<&str>> = terms.iter().map(|t| vec![t.as_str()]).collect::<Vec<_>>();
-        // single segment so we have a single dictionary of 32 terms.
-        let index = get_test_index_from_terms(true, &term_refs)?;
-
-        let agg_req: Aggregations = serde_json::from_value(json!({
-            "cardinality": {
-                "cardinality": { "field": "string_id" }
-            },
-        }))
-        .unwrap();
-
-        let res = exec_request(agg_req, &index)?;
-        assert_eq!(res["cardinality"]["value"], 32.0);
-        Ok(())
-    }
-
-    /// `BitSet` path with a `missing` parameter: the column-level missing
-    /// sentinel (`column.max_value() + 1`) flows into the bitset, the
-    /// dict lookup filter at finalization drops it, and the missing
-    /// coupon is applied separately.
-    #[test]
-    fn cardinality_aggregation_test_str_bitset_with_missing() {
-        let mut schema_builder = Schema::builder();
-        let name_field = schema_builder.add_text_field("name", STRING | FAST);
-        let index = Index::create_in_ram(schema_builder.build());
-        let mut writer = index.writer_for_tests().unwrap();
-        for i in 0..16 {
-            let term = format!("t{i:02}");
-            writer.add_document(doc!(name_field => term)).unwrap();
-        }
-        // One empty doc, exercising the missing sentinel.
-        writer.add_document(doc!()).unwrap();
-        writer.commit().unwrap();
-
-        let agg_req: Aggregations = serde_json::from_value(json!({
-            "cardinality": {
-                "cardinality": {
-                    "field": "name",
-                    "missing": "MISSING_SENTINEL_KEY",
-                }
-            },
-        }))
-        .unwrap();
-
-        let res = exec_request(agg_req, &index).unwrap();
-        // 16 distinct real terms + 1 distinct "missing" value = 17.
-        assert_eq!(res["cardinality"]["value"], 17.0);
-    }
-
-    /// Unit-test the PagedBitset itself: cross-page inserts produce sorted
-    /// iteration, len() matches the inserted set, and duplicates are
-    /// idempotent.
-    #[test]
-    fn paged_bitset_basic() {
-        use super::PagedBitset;
-        // Span several pages: BITSET_PAGE_BITS = 1024, so ords > 1024 land
-        // on the second page, > 2048 on the third, etc.
-        let ords = [0u64, 1, 63, 64, 1023, 1024, 1025, 4096, 4097, 9999, 10_000];
-        let max_ord = *ords.iter().max().unwrap();
-        let mut bitset = PagedBitset::with_max_term_ord(max_ord);
-        for &ord in &ords {
-            bitset.insert(ord);
-            // Idempotent: inserting again must not increase count.
-            bitset.insert(ord);
-        }
-        assert_eq!(bitset.len(), ords.len() as u64);
-        let collected: Vec<u64> = bitset.iter_sorted().collect();
-        let mut expected: Vec<u64> = ords.to_vec();
-        expected.sort_unstable();
-        assert_eq!(collected, expected);
-    }
-
-    /// Unit-test `TermOrdSet`: starts Sparse, promotes to Dense on
-    /// `maybe_compact` once the density threshold is crossed, and
-    /// `iter_ords()` yields the same set in either state. Ords spanning
-    /// multiple paged-bitset pages exercise the Dense iter ordering.
-    #[test]
-    fn term_ord_set_promotes_on_maybe_compact() {
-        use super::{TermOrdAccumulator, TermOrdSet, PROMOTION_RATIO};
-        // Pick max so promotion needs few inserts: len * RATIO > max with
-        // RATIO=32 and max=64 trips at len=3 (3*32=96 > 64).
-        let max_term_ord = 64u64;
-        let mut set = <TermOrdSet as TermOrdAccumulator>::new(max_term_ord);
-        // Two inserts: should stay Sparse after maybe_compact (2 * RATIO = 64, not > 64).
-        set.insert(0);
-        set.insert(7);
-        set.maybe_compact();
-        assert_eq!(set.len(), 2);
-
-        // Third insert promotes on next maybe_compact.
-        set.insert(20);
-        assert_eq!(set.len(), 3);
-        // Sanity check: at len=3, 3 * PROMOTION_RATIO = 96 > 64.
-        assert!(3u64 * PROMOTION_RATIO > max_term_ord);
-        set.maybe_compact();
-
-        // Post-promotion: extending continues to work.
-        set.insert(15);
-        set.insert(15); // dup
-        assert_eq!(set.len(), 4);
-
-        let mut collected: Vec<u64> = set.iter_ords().collect();
-        collected.sort_unstable();
-        assert_eq!(collected, vec![0, 7, 15, 20]);
-    }
-
-    /// Unit-test the `BitSet` impl of `TermOrdAccumulator`: insert,
-    /// dedup, and iter_ords order.
-    #[test]
-    fn bitset_accumulator_basic() {
-        use common::BitSet;
-
-        use super::TermOrdAccumulator;
-        let mut set = <BitSet as TermOrdAccumulator>::new(255);
-        for ord in [0u64, 1, 63, 64, 65, 128, 200, 200, 0] {
-            <BitSet as TermOrdAccumulator>::insert(&mut set, ord);
-        }
-        assert_eq!(<BitSet as TermOrdAccumulator>::len(&set), 7);
-        let collected: Vec<u64> = set.iter_ords().collect();
-        assert_eq!(collected, vec![0, 1, 63, 64, 65, 128, 200]);
-    }
-
-    #[test]
-    fn cardinality_aggregation_u64() -> crate::Result<()> {
-        let mut schema_builder = Schema::builder();
-        let id_field = schema_builder.add_u64_field("id", FAST);
-        let index = Index::create_in_ram(schema_builder.build());
-        {
-            let mut writer = index.writer_for_tests()?;
-            writer.add_document(doc!(id_field => 1u64))?;
-            writer.add_document(doc!(id_field => 2u64))?;
-            writer.add_document(doc!(id_field => 3u64))?;
-            writer.add_document(doc!())?;
-            writer.commit()?;
-        }
-
-        let agg_req: Aggregations = serde_json::from_value(json!({
-            "cardinality": {
-                "cardinality": {
-                    "field": "id",
-                    "missing": 0u64
-                },
-            }
-        }))
-        .unwrap();
-
-        let res = exec_request(agg_req, &index)?;
-        assert_eq!(res["cardinality"]["value"], 4.0);
-
-        Ok(())
-    }
-
-    #[test]
-    fn cardinality_aggregation_ip_addr() -> crate::Result<()> {
-        let mut schema_builder = Schema::builder();
-        let field = schema_builder.add_ip_addr_field("ip_field", FAST);
-        let index = Index::create_in_ram(schema_builder.build());
-        {
-            let mut writer = index.writer_for_tests()?;
-            // IpV6 loopback
-            writer.add_document(doc!(field=>IpAddr::from_str("::1").unwrap().into_ipv6_addr()))?;
-            writer.add_document(doc!(field=>IpAddr::from_str("::1").unwrap().into_ipv6_addr()))?;
-            // IpV4
-            writer.add_document(
-                doc!(field=>IpAddr::from_str("127.0.0.1").unwrap().into_ipv6_addr()),
-            )?;
-            writer.commit()?;
-        }
-
-        let agg_req: Aggregations = serde_json::from_value(json!({
-            "cardinality": {
-                "cardinality": {
-                    "field": "ip_field"
-                },
-            }
-        }))
-        .unwrap();
-
-        let res = exec_request(agg_req, &index)?;
-        assert_eq!(res["cardinality"]["value"], 2.0);
-
-        Ok(())
-    }
-
-    #[test]
-    fn cardinality_aggregation_bytes_excluded_from_accessors() -> crate::Result<()> {
-        // `Bytes` columns are opened as raw per-segment dictionary ordinals (like `Str`), but
-        // unlike `Str`, cardinality has no dictionary-resolution path for them: it would hash
-        // the raw ordinal directly, which are segment dependant. Ignore bytes values instead of
-        // counting them wrong.
-        let mut schema_builder = Schema::builder();
-        let field = schema_builder.add_bytes_field("raw", FAST);
-        let index = Index::create_in_ram(schema_builder.build());
-        {
-            let mut writer = index.writer_for_tests()?;
-            writer.add_document(doc!(field => vec![1u8]))?;
-            writer.add_document(doc!(field => vec![2u8]))?;
-            writer.commit()?;
-            writer.add_document(doc!(field => vec![3u8]))?;
-            writer.add_document(doc!(field => vec![4u8]))?;
-            writer.commit()?;
-        }
-
-        let agg_req: Aggregations = serde_json::from_value(json!({
-            "cardinality": {
-                "cardinality": {
-                    "field": "raw"
-                },
-            }
-        }))
-        .unwrap();
-
-        let res = exec_request(agg_req, &index)?;
-        assert_eq!(res["cardinality"]["value"], 0.0);
-
-        Ok(())
-    }
-
-    #[test]
-    fn cardinality_aggregation_json() -> crate::Result<()> {
-        let mut schema_builder = Schema::builder();
-        let field = schema_builder.add_json_field("json", FAST);
-        let index = Index::create_in_ram(schema_builder.build());
-        {
-            let mut writer = index.writer_for_tests()?;
-            writer.add_document(doc!(field => json!({"value": false})))?;
-            writer.add_document(doc!(field => json!({"value": true})))?;
-            writer.add_document(doc!(field => json!({"value": i64::from_u64(0u64)})))?;
-            writer.add_document(doc!(field => json!({"value": i64::from_u64(1u64)})))?;
-            writer.commit()?;
-        }
-
-        let agg_req: Aggregations = serde_json::from_value(json!({
-            "cardinality": {
-                "cardinality": {
-                    "field": "json.value"
-                },
-            }
-        }))
-        .unwrap();
-
-        let res = exec_request(agg_req, &index)?;
-        assert_eq!(res["cardinality"]["value"], 4.0);
-
-        Ok(())
-    }
-
-    /// A JSON path that resolves to both a Str column and a numeric column
-    /// produces two collector instances per segment — one with `Str` buckets
-    /// and one with `Numeric` buckets. Their `IntermediateMetricResult`s must
-    /// merge into the union cardinality.
-    #[test]
-    fn cardinality_aggregation_json_str_and_numeric() -> crate::Result<()> {
-        let mut schema_builder = Schema::builder();
-        let field = schema_builder.add_json_field("json", FAST);
-        let index = Index::create_in_ram(schema_builder.build());
-        {
-            let mut writer = index.writer_for_tests()?;
-            writer.add_document(doc!(field => json!({"value": "hello"})))?;
-            writer.add_document(doc!(field => json!({"value": "world"})))?;
-            writer.add_document(doc!(field => json!({"value": "hello"})))?; // dup str
-            writer.add_document(doc!(field => json!({"value": i64::from_u64(7u64)})))?;
-            writer.add_document(doc!(field => json!({"value": i64::from_u64(42u64)})))?;
-            writer.add_document(doc!(field => json!({"value": i64::from_u64(7u64)})))?; // dup num
-            writer.commit()?;
-        }
-
-        let agg_req: Aggregations = serde_json::from_value(json!({
-            "cardinality": {
-                "cardinality": {
-                    "field": "json.value"
-                },
-            }
-        }))
-        .unwrap();
-
-        let res = exec_request(agg_req, &index)?;
-        // 4 distinct values: "hello", "world", 7, 42.
-        assert_eq!(res["cardinality"]["value"], 4.0);
-
-        Ok(())
-    }
-
-    #[test]
-    fn cardinality_collector_serde_roundtrip() {
-        use super::CardinalityCollector;
-
-        let mut collector = CardinalityCollector::default();
-        collector.insert("hello");
-        collector.insert("world");
-        collector.insert("hello"); // duplicate
-
-        let serialized = serde_json::to_vec(&collector).unwrap();
-        let deserialized: CardinalityCollector = serde_json::from_slice(&serialized).unwrap();
-
-        let original_estimate = collector.finalize().unwrap();
-        let roundtrip_estimate = deserialized.finalize().unwrap();
-        assert_eq!(original_estimate, roundtrip_estimate);
-        assert_eq!(original_estimate, 2.0);
-    }
-
-    #[test]
-    fn cardinality_collector_merge() {
-        use super::CardinalityCollector;
-
-        let mut left = CardinalityCollector::default();
-        left.insert("a");
-        left.insert("b");
-
-        let mut right = CardinalityCollector::default();
-        right.insert("b");
-        right.insert("c");
-
-        left.merge_fruits(right).unwrap();
-        let estimate = left.finalize().unwrap();
-        assert_eq!(estimate, 3.0);
-    }
-
-    /// Verifies that merging two small sketches (both in List/Set coupon mode)
-    /// produces an exact result — i.e. the HllUnion does not unnecessarily
-    /// promote to the full HLL array when the combined cardinality is small.
-    #[test]
-    fn cardinality_collector_merge_stays_exact_for_small_sets() {
-        use super::CardinalityCollector;
-
-        let mut left = CardinalityCollector::default();
-        for i in 0u64..50 {
-            left.insert(i);
-        }
-
-        let mut right = CardinalityCollector::default();
-        for i in 30u64..100 {
-            right.insert(i);
-        }
-
-        left.merge_fruits(right).unwrap();
-        let estimate = left.finalize().unwrap();
-        // 100 distinct values (0..100). Both sketches are in Set mode (< 192 coupons),
-        // so the union should stay in coupon mode and give an exact count.
-        assert_eq!(estimate, 100.0);
-    }
-
-    #[test]
-    fn cardinality_collector_serialize_deserialize_binary() {
-        use datasketches::hll::HllSketch;
-
-        use super::CardinalityCollector;
-
-        let mut collector = CardinalityCollector::default();
-        collector.insert("apple");
-        collector.insert("banana");
-        collector.insert("cherry");
-
-        let bytes = collector.to_sketch_bytes();
-        let deserialized = HllSketch::deserialize(&bytes).unwrap();
-        assert!((deserialized.estimate() - 3.0).abs() < 0.01);
-    }
-
-    /// Tests that the `missing` parameter correctly counts a single empty document
-    /// for both u64 and str columns.
-    #[test]
-    fn cardinality_aggregation_missing_value_single_empty_doc() {
-        let mut schema_builder = Schema::builder();
-        let id_field = schema_builder.add_u64_field("id", FAST);
-        let name_field = schema_builder.add_text_field("name", STRING | FAST);
-        let index = Index::create_in_ram(schema_builder.build());
-        let mut writer = index.writer_for_tests().unwrap();
-        writer
-            .add_document(doc!(id_field=>1u64,name_field=>"some_name"))
-            .unwrap();
-        writer.add_document(doc!()).unwrap();
-        writer.commit().unwrap();
-
-        {
-            // int colum with missing value non redundant
-            let agg_req: Aggregations = serde_json::from_value(json!({
-                "cardinality": {
-                    "cardinality": {
-                        "field": "id",
-                        "missing": 42u64
-                    },
-                }
-            }))
-            .unwrap();
-            let res = exec_request(agg_req, &index).unwrap();
-            assert_eq!(res["cardinality"]["value"], 2.0);
-        }
-
-        {
-            // int colum with missing value redundant
-            let agg_req: Aggregations = serde_json::from_value(json!({
-                "cardinality": {
-                    "cardinality": {
-                        "field": "id",
-                        "missing": 1u64
-                    },
-                }
-            }))
-            .unwrap();
-            let res = exec_request(agg_req, &index).unwrap();
-            assert_eq!(res["cardinality"]["value"], 1.0);
-        }
-
-        {
-            // str colum with missing value non redundant
-            // With more than one segment, this is not well handled.
-            let agg_req: Aggregations = serde_json::from_value(json!({
-                "cardinality": {
-                    "cardinality": {
-                        "field": "name",
-                        "missing": "other_name"
-                    },
-                }
-            }))
-            .unwrap();
-            let res = exec_request(agg_req, &index).unwrap();
-            assert_eq!(res["cardinality"]["value"], 2.0);
-        }
-
-        {
-            // str colum with missing value redundant
-            let agg_req: Aggregations = serde_json::from_value(json!({
-                "cardinality": {
-                    "cardinality": {
-                        "field": "name",
-                        "missing": "some_name"
-                    },
-                }
-            }))
-            .unwrap();
-            let res = exec_request(agg_req, &index).unwrap();
-            assert_eq!(res["cardinality"]["value"], 1.0);
-        }
-
-        {
-            // str column with missing value with a number type.
-            let agg_req: Aggregations = serde_json::from_value(json!({
-                "cardinality": {
-                    "cardinality": {
-                        "field": "name",
-                        "missing": 3,
-                    },
-                }
-            }))
-            .unwrap();
-            let res = exec_request(agg_req, &index).unwrap();
-            assert_eq!(res["cardinality"]["value"], 2.0);
-        }
-    }
-
-    #[test]
-    fn cardinality_collector_salt_differentiates_types() {
-        use super::CardinalityCollector;
-
-        // Without salt, same u64 value from different column types would collide
-        let mut collector_bool = CardinalityCollector::new(5); // e.g. ColumnType::Bool
-        collector_bool.insert(0u64); // false
-        collector_bool.insert(1u64); // true
-
-        let mut collector_i64 = CardinalityCollector::new(2); // e.g. ColumnType::I64
-        collector_i64.insert(0u64);
-        collector_i64.insert(1u64);
-
-        // Merge them
-        collector_bool.merge_fruits(collector_i64).unwrap();
-        let estimate = collector_bool.finalize().unwrap();
-        // Should be 4 because salt makes (5, 0) != (2, 0) and (5, 1) != (2, 1)
-        assert_eq!(estimate, 4.0);
-    }
-}
diff --git a/src/aggregation/metric/cardinality/mod.rs b/src/aggregation/metric/cardinality/mod.rs
new file mode 100644
index 000000000..8440cd948
--- /dev/null
+++ b/src/aggregation/metric/cardinality/mod.rs
@@ -0,0 +1,614 @@
+//! Cardinality aggregation.
+//!
+//!   * [`str_collector`] holds the `ColumnType::Str` segment collector, which accumulates term
+//!     ordinals and resolves them to HLL coupons at finalization.
+//!   * [`numeric_collector`] holds the segment collector for every other column type, which feeds
+//!     the HLL sketch directly during collection.
+//!
+//! Both segment collectors converge on the same
+//! [`IntermediateMetricResult::Cardinality`] payload, so results coming from a
+//! str column and from a numeric column (e.g. a JSON path that resolves to
+//! both) merge through the single [`CardinalityCollector::merge_fruits`]
+//! implementation here.
+
+mod numeric_collector;
+mod str_collector;
+mod term_ord_accumulator;
+
+use std::hash::Hash;
+use std::sync::Arc;
+
+use columnar::{ColumnType, StrColumn};
+use common::BitSet;
+use datasketches::hll::{Coupon, HllSketch, HllType, HllUnion};
+pub(crate) use numeric_collector::SegmentNumericCardinalityCollector;
+use serde::{Deserialize, Deserializer, Serialize, Serializer};
+pub(crate) use str_collector::SegmentStrCardinalityCollector;
+pub(crate) use term_ord_accumulator::{TermOrdSet, BITSET_MAX_TERM_ORD};
+
+use crate::aggregation::agg_data::{AggRefNode, AggregationsSegmentCtx};
+use crate::aggregation::segment_agg_result::SegmentAggregationCollector;
+use crate::aggregation::value_source::ValueSource;
+use crate::aggregation::*;
+
+/// Log2 of the number of registers for the HLL sketch.
+/// 2^11 = 2048 registers, giving ~2.3% relative error and ~1KB per sketch (Hll4).
+const LG_K: u8 = 11;
+
+/// # Cardinality
+///
+/// The cardinality aggregation allows for computing an estimate
+/// of the number of different values in a data set based on the
+/// Apache DataSketches HyperLogLog algorithm. This is particularly useful for
+/// understanding the uniqueness of values in a large dataset where counting
+/// each unique value individually would be computationally expensive.
+///
+/// For example, you might use a cardinality aggregation to estimate the number
+/// of unique visitors to a website by aggregating on a field that contains
+/// user IDs or session IDs.
+///
+/// To use the cardinality aggregation, you'll need to provide a field to
+/// aggregate on. The following example demonstrates a request for the cardinality
+/// of the "user_id" field:
+///
+/// ```JSON
+/// {
+///     "cardinality": {
+///         "field": "user_id"
+///     }
+/// }
+/// ```
+///
+/// This request will return an estimate of the number of unique values in the
+/// "user_id" field.
+///
+/// ## Missing Values
+///
+/// The `missing` parameter defines how documents that are missing a value should be treated.
+/// By default, documents without a value for the specified field are ignored. However, you can
+/// specify a default value for these documents using the `missing` parameter. This can be useful
+/// when you want to include documents with missing values in the aggregation.
+///
+/// For example, the following request treats documents with missing values in the "user_id"
+/// field as if they had a value of "unknown":
+///
+/// ```JSON
+/// {
+///     "cardinality": {
+///         "field": "user_id",
+///         "missing": "unknown"
+///     }
+/// }
+/// ```
+///
+/// # Estimation Accuracy
+///
+/// The cardinality aggregation provides an approximate count, which is usually
+/// accurate within a small error range. This trade-off allows for efficient
+/// computation even on very large datasets.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct CardinalityAggregationReq {
+    /// The field name to compute the percentiles on.
+    pub field: String,
+    /// The missing parameter defines how documents that are missing a value should be treated.
+    /// By default they will be ignored but it is also possible to treat them as if they had a
+    /// value. Examples in JSON format:
+    /// { "field": "my_numbers", "missing": "10.0" }
+    #[serde(skip_serializing_if = "Option::is_none", default)]
+    pub missing: Option<Key>,
+}
+
+/// Contains all information required by the segment cardinality collectors to perform the
+/// cardinality aggregation on a segment.
+pub(crate) struct CardinalityAggReqData {
+    /// The column accessor to access the fast field values.
+    pub(crate) accessor: Arc<dyn ValueSource>,
+    /// The string dictionary column if the field is of type string.
+    pub(crate) str_dict_column: Option<StrColumn>,
+    /// The missing value normalized to the internal u64 representation of the field type.
+    pub(crate) missing_value_for_accessor: Option<u64>,
+    /// The name of the aggregation.
+    pub(crate) name: String,
+    /// The aggregation request.
+    pub(crate) req: CardinalityAggregationReq,
+}
+
+impl CardinalityAggReqData {
+    /// Estimate the memory consumption of this struct in bytes.
+    pub fn get_memory_consumption(&self) -> usize {
+        std::mem::size_of::<Self>()
+    }
+}
+
+impl CardinalityAggregationReq {
+    /// Creates a new [`CardinalityAggregationReq`] instance from a field name.
+    pub fn from_field_name(field_name: String) -> Self {
+        Self {
+            field: field_name,
+            missing: None,
+        }
+    }
+    /// Returns the field name the aggregation is computed on.
+    pub fn field_name(&self) -> &str {
+        &self.field
+    }
+}
+
+#[derive(Clone, Debug)]
+/// The cardinality collector used during segment collection and for merging results.
+/// Uses Apache DataSketches HLL (lg_k=11, Hll4) for compact binary serialization
+/// and cross-language compatibility (e.g. Java `datasketches` library).
+pub struct CardinalityCollector {
+    sketch: HllSketch,
+    /// Salt derived from `ColumnType`, used to differentiate values of different column types
+    /// that map to the same u64 (e.g. bool `false` = 0 vs i64 `0`).
+    /// Not serialized — only needed during insertion, not after sketch registers are populated.
+    salt: u8,
+}
+
+impl Default for CardinalityCollector {
+    fn default() -> Self {
+        Self::new(0)
+    }
+}
+
+impl PartialEq for CardinalityCollector {
+    fn eq(&self, _other: &Self) -> bool {
+        false
+    }
+}
+
+impl Serialize for CardinalityCollector {
+    fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
+        let bytes = self.sketch.serialize();
+        serializer.serialize_bytes(&bytes)
+    }
+}
+
+impl<'de> Deserialize<'de> for CardinalityCollector {
+    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
+        let bytes: Vec<u8> = Deserialize::deserialize(deserializer)?;
+        let sketch = HllSketch::deserialize(&bytes).map_err(serde::de::Error::custom)?;
+        Ok(Self { sketch, salt: 0 })
+    }
+}
+
+impl CardinalityCollector {
+    fn new(salt: u8) -> Self {
+        Self {
+            sketch: HllSketch::new(LG_K, HllType::Hll8)
+                .expect("LG_K is within the supported range"),
+            salt,
+        }
+    }
+
+    /// Insert a value into the HLL sketch, salted by the column type.
+    /// The salt ensures that identical u64 values from different column types
+    /// (e.g. bool `false` vs i64 `0`) are counted as distinct.
+    fn insert(&mut self, value: impl Hash) {
+        self.sketch.update((self.salt, value));
+    }
+
+    fn insert_coupon(&mut self, coupon: Coupon) {
+        self.sketch.update_with_coupon(coupon);
+    }
+
+    /// Compute the final cardinality estimate.
+    pub fn finalize(self) -> Option<f64> {
+        Some(self.sketch.estimate().trunc())
+    }
+
+    /// Serialize the HLL sketch to its compact binary representation.
+    /// The format is cross-language compatible with Apache DataSketches (Java, C++, Python).
+    pub fn to_sketch_bytes(&self) -> Vec<u8> {
+        self.sketch.serialize()
+    }
+
+    pub(crate) fn merge_fruits(&mut self, right: CardinalityCollector) -> crate::Result<()> {
+        let mut union = HllUnion::new(LG_K).expect("LG_K is within the supported range");
+        union.update(&self.sketch);
+        union.update(&right.sketch);
+        self.sketch = union.to_sketch(HllType::Hll8);
+        Ok(())
+    }
+}
+
+/// Builds the segment collector for a cardinality aggregation.
+///
+/// str and non-str columns use two entirely different collectors: str
+/// accumulates term ordinals and resolves them into HLL coupons at
+/// finalization, non-str feeds the HLL sketch directly. Both produce the same
+/// [`IntermediateMetricResult::Cardinality`], so they merge uniformly.
+pub(crate) fn build_segment_cardinality_collector(
+    req: &mut AggregationsSegmentCtx,
+    node: &AggRefNode,
+) -> crate::Result<Box<dyn SegmentAggregationCollector>> {
+    let req_data = req.get_cardinality_req_data(node.idx_in_req_data);
+    if req_data.accessor.column_type() != ColumnType::Str {
+        return Ok(Box::new(SegmentNumericCardinalityCollector::from_req(
+            node.idx_in_req_data,
+            req_data.accessor.clone(),
+            req_data.missing_value_for_accessor,
+        )?));
+    }
+    // For str columns, we need to collect the set of term ordinals encounterred.
+    // We choose a different representation depending on the number of maximum
+    // number of terms.
+    //   * small (< BITSET_MAX_TERM_ORD): `BitSet`, pre-allocated.
+    //   * large: `TermOrdSet` (sparse HashSet that promotes to a paged bitset).
+    let Some(column) = req_data.accessor.as_column() else {
+        return Err(crate::TantivyError::InvalidArgument(
+            "cardinality over str virtual columns is not supported yet".to_string(),
+        ));
+    };
+    let max_term_ord_inclusive = column.max_value();
+    if max_term_ord_inclusive < BITSET_MAX_TERM_ORD {
+        Ok(Box::new(
+            SegmentStrCardinalityCollector::<BitSet>::from_req(
+                node.idx_in_req_data,
+                req_data.accessor.clone(),
+                req_data.missing_value_for_accessor,
+                max_term_ord_inclusive,
+            ),
+        ))
+    } else {
+        Ok(Box::new(
+            SegmentStrCardinalityCollector::<TermOrdSet>::from_req(
+                node.idx_in_req_data,
+                req_data.accessor.clone(),
+                req_data.missing_value_for_accessor,
+                max_term_ord_inclusive,
+            ),
+        ))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use columnar::MonotonicallyMappableToU64;
+
+    use crate::aggregation::agg_req::Aggregations;
+    use crate::aggregation::tests::{exec_request, get_test_index_from_terms};
+    use crate::schema::{Schema, FAST, STRING};
+    use crate::Index;
+
+    #[test]
+    fn cardinality_aggregation_test_empty_index() -> crate::Result<()> {
+        let values = vec![];
+        let index = get_test_index_from_terms(false, &values)?;
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "cardinality": {
+                "cardinality": {
+                    "field": "string_id",
+                }
+            },
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["cardinality"]["value"], 0.0);
+
+        Ok(())
+    }
+
+    #[test]
+    fn cardinality_aggregation_test_single_segment() -> crate::Result<()> {
+        cardinality_aggregation_test_merge_segment(true)
+    }
+    #[test]
+    fn cardinality_aggregation_test() -> crate::Result<()> {
+        cardinality_aggregation_test_merge_segment(false)
+    }
+    fn cardinality_aggregation_test_merge_segment(merge_segments: bool) -> crate::Result<()> {
+        let segment_and_terms = vec![
+            vec!["terma"],
+            vec!["termb"],
+            vec!["termc"],
+            vec!["terma"],
+            vec!["terma"],
+            vec!["terma"],
+            vec!["termb"],
+            vec!["terma"],
+        ];
+        let index = get_test_index_from_terms(merge_segments, &segment_and_terms)?;
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "cardinality": {
+                "cardinality": {
+                    "field": "string_id",
+                }
+            },
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["cardinality"]["value"], 3.0);
+
+        Ok(())
+    }
+
+    #[test]
+    fn cardinality_aggregation_bytes_excluded_from_accessors() -> crate::Result<()> {
+        // `Bytes` columns are opened as raw per-segment dictionary ordinals (like `Str`), but
+        // unlike `Str`, cardinality has no dictionary-resolution path for them: it would hash
+        // the raw ordinal directly, which are segment dependant. Ignore bytes values instead of
+        // counting them wrong.
+        let mut schema_builder = Schema::builder();
+        let field = schema_builder.add_bytes_field("raw", FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        {
+            let mut writer = index.writer_for_tests()?;
+            writer.add_document(doc!(field => vec![1u8]))?;
+            writer.add_document(doc!(field => vec![2u8]))?;
+            writer.commit()?;
+            writer.add_document(doc!(field => vec![3u8]))?;
+            writer.add_document(doc!(field => vec![4u8]))?;
+            writer.commit()?;
+        }
+
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "cardinality": {
+                "cardinality": {
+                    "field": "raw"
+                },
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["cardinality"]["value"], 0.0);
+
+        Ok(())
+    }
+
+    #[test]
+    fn cardinality_aggregation_json() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let field = schema_builder.add_json_field("json", FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        {
+            let mut writer = index.writer_for_tests()?;
+            writer.add_document(doc!(field => json!({"value": false})))?;
+            writer.add_document(doc!(field => json!({"value": true})))?;
+            writer.add_document(doc!(field => json!({"value": i64::from_u64(0u64)})))?;
+            writer.add_document(doc!(field => json!({"value": i64::from_u64(1u64)})))?;
+            writer.commit()?;
+        }
+
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "cardinality": {
+                "cardinality": {
+                    "field": "json.value"
+                },
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["cardinality"]["value"], 4.0);
+
+        Ok(())
+    }
+
+    /// A JSON path that resolves to both a Str column and a numeric column
+    /// produces two collector instances per segment — one with `Str` buckets
+    /// and one with `Numeric` buckets. Their `IntermediateMetricResult`s must
+    /// merge into the union cardinality.
+    #[test]
+    fn cardinality_aggregation_json_str_and_numeric() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let field = schema_builder.add_json_field("json", FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        {
+            let mut writer = index.writer_for_tests()?;
+            writer.add_document(doc!(field => json!({"value": "hello"})))?;
+            writer.add_document(doc!(field => json!({"value": "world"})))?;
+            writer.add_document(doc!(field => json!({"value": "hello"})))?; // dup str
+            writer.add_document(doc!(field => json!({"value": i64::from_u64(7u64)})))?;
+            writer.add_document(doc!(field => json!({"value": i64::from_u64(42u64)})))?;
+            writer.add_document(doc!(field => json!({"value": i64::from_u64(7u64)})))?; // dup num
+            writer.commit()?;
+        }
+
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "cardinality": {
+                "cardinality": {
+                    "field": "json.value"
+                },
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        // 4 distinct values: "hello", "world", 7, 42.
+        assert_eq!(res["cardinality"]["value"], 4.0);
+
+        Ok(())
+    }
+
+    #[test]
+    fn cardinality_collector_serde_roundtrip() {
+        use super::CardinalityCollector;
+
+        let mut collector = CardinalityCollector::default();
+        collector.insert("hello");
+        collector.insert("world");
+        collector.insert("hello"); // duplicate
+
+        let serialized = serde_json::to_vec(&collector).unwrap();
+        let deserialized: CardinalityCollector = serde_json::from_slice(&serialized).unwrap();
+
+        let original_estimate = collector.finalize().unwrap();
+        let roundtrip_estimate = deserialized.finalize().unwrap();
+        assert_eq!(original_estimate, roundtrip_estimate);
+        assert_eq!(original_estimate, 2.0);
+    }
+
+    #[test]
+    fn cardinality_collector_merge() {
+        use super::CardinalityCollector;
+
+        let mut left = CardinalityCollector::default();
+        left.insert("a");
+        left.insert("b");
+
+        let mut right = CardinalityCollector::default();
+        right.insert("b");
+        right.insert("c");
+
+        left.merge_fruits(right).unwrap();
+        let estimate = left.finalize().unwrap();
+        assert_eq!(estimate, 3.0);
+    }
+
+    /// Verifies that merging two small sketches (both in List/Set coupon mode)
+    /// produces an exact result — i.e. the HllUnion does not unnecessarily
+    /// promote to the full HLL array when the combined cardinality is small.
+    #[test]
+    fn cardinality_collector_merge_stays_exact_for_small_sets() {
+        use super::CardinalityCollector;
+
+        let mut left = CardinalityCollector::default();
+        for i in 0u64..50 {
+            left.insert(i);
+        }
+
+        let mut right = CardinalityCollector::default();
+        for i in 30u64..100 {
+            right.insert(i);
+        }
+
+        left.merge_fruits(right).unwrap();
+        let estimate = left.finalize().unwrap();
+        // 100 distinct values (0..100). Both sketches are in Set mode (< 192 coupons),
+        // so the union should stay in coupon mode and give an exact count.
+        assert_eq!(estimate, 100.0);
+    }
+
+    #[test]
+    fn cardinality_collector_serialize_deserialize_binary() {
+        use datasketches::hll::HllSketch;
+
+        use super::CardinalityCollector;
+
+        let mut collector = CardinalityCollector::default();
+        collector.insert("apple");
+        collector.insert("banana");
+        collector.insert("cherry");
+
+        let bytes = collector.to_sketch_bytes();
+        let deserialized = HllSketch::deserialize(&bytes).unwrap();
+        assert!((deserialized.estimate() - 3.0).abs() < 0.01);
+    }
+
+    /// Tests that the `missing` parameter correctly counts a single empty document
+    /// for both u64 and str columns.
+    #[test]
+    fn cardinality_aggregation_missing_value_single_empty_doc() {
+        let mut schema_builder = Schema::builder();
+        let id_field = schema_builder.add_u64_field("id", FAST);
+        let name_field = schema_builder.add_text_field("name", STRING | FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut writer = index.writer_for_tests().unwrap();
+        writer
+            .add_document(doc!(id_field=>1u64,name_field=>"some_name"))
+            .unwrap();
+        writer.add_document(doc!()).unwrap();
+        writer.commit().unwrap();
+
+        {
+            // int colum with missing value non redundant
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "cardinality": {
+                    "cardinality": {
+                        "field": "id",
+                        "missing": 42u64
+                    },
+                }
+            }))
+            .unwrap();
+            let res = exec_request(agg_req, &index).unwrap();
+            assert_eq!(res["cardinality"]["value"], 2.0);
+        }
+
+        {
+            // int colum with missing value redundant
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "cardinality": {
+                    "cardinality": {
+                        "field": "id",
+                        "missing": 1u64
+                    },
+                }
+            }))
+            .unwrap();
+            let res = exec_request(agg_req, &index).unwrap();
+            assert_eq!(res["cardinality"]["value"], 1.0);
+        }
+
+        {
+            // str colum with missing value non redundant
+            // With more than one segment, this is not well handled.
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "cardinality": {
+                    "cardinality": {
+                        "field": "name",
+                        "missing": "other_name"
+                    },
+                }
+            }))
+            .unwrap();
+            let res = exec_request(agg_req, &index).unwrap();
+            assert_eq!(res["cardinality"]["value"], 2.0);
+        }
+
+        {
+            // str colum with missing value redundant
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "cardinality": {
+                    "cardinality": {
+                        "field": "name",
+                        "missing": "some_name"
+                    },
+                }
+            }))
+            .unwrap();
+            let res = exec_request(agg_req, &index).unwrap();
+            assert_eq!(res["cardinality"]["value"], 1.0);
+        }
+
+        {
+            // str column with missing value with a number type.
+            let agg_req: Aggregations = serde_json::from_value(json!({
+                "cardinality": {
+                    "cardinality": {
+                        "field": "name",
+                        "missing": 3,
+                    },
+                }
+            }))
+            .unwrap();
+            let res = exec_request(agg_req, &index).unwrap();
+            assert_eq!(res["cardinality"]["value"], 2.0);
+        }
+    }
+
+    #[test]
+    fn cardinality_collector_salt_differentiates_types() {
+        use super::CardinalityCollector;
+
+        // Without salt, same u64 value from different column types would collide
+        let mut collector_bool = CardinalityCollector::new(5); // e.g. ColumnType::Bool
+        collector_bool.insert(0u64); // false
+        collector_bool.insert(1u64); // true
+
+        let mut collector_i64 = CardinalityCollector::new(2); // e.g. ColumnType::I64
+        collector_i64.insert(0u64);
+        collector_i64.insert(1u64);
+
+        // Merge them
+        collector_bool.merge_fruits(collector_i64).unwrap();
+        let estimate = collector_bool.finalize().unwrap();
+        // Should be 4 because salt makes (5, 0) != (2, 0) and (5, 1) != (2, 1)
+        assert_eq!(estimate, 4.0);
+    }
+}
diff --git a/src/aggregation/metric/cardinality/numeric_collector.rs b/src/aggregation/metric/cardinality/numeric_collector.rs
new file mode 100644
index 000000000..d6b3fa7bc
--- /dev/null
+++ b/src/aggregation/metric/cardinality/numeric_collector.rs
@@ -0,0 +1,264 @@
+//! Segment collector for `cardinality` over any non-str column
+//! (numeric, bool, date, IpAddr).
+//!
+//! Unlike the str case there is no dictionary to resolve, so values go
+//! straight into a per-bucket HLL sketch during collection. The produced
+//! [`CardinalityCollector`] is the same type the str collector produces, so
+//! intermediate merging stays in the parent module.
+
+use std::fmt::Debug;
+use std::sync::Arc;
+
+use columnar::column_values::CompactSpaceU64Accessor;
+use columnar::ColumnType;
+
+use super::CardinalityCollector;
+use crate::aggregation::agg_data::AggregationsSegmentCtx;
+use crate::aggregation::intermediate_agg_result::{
+    IntermediateAggregationResult, IntermediateAggregationResults, IntermediateMetricResult,
+};
+use crate::aggregation::segment_agg_result::SegmentAggregationCollector;
+use crate::aggregation::value_source::ValueSource;
+use crate::aggregation::*;
+use crate::TantivyError;
+
+/// Segment collector for `cardinality` over any non-str column
+/// (numeric, bool, date, IpAddr).
+///
+/// Hidden contract: `column_type` must not be `ColumnType::Str`. Values are
+/// inserted into the HLL sketch during collection, so the sketch of a bucket
+/// is already complete when the bucket is finalized.
+pub(crate) struct SegmentNumericCardinalityCollector {
+    /// Buckets are Some(_) until they get consumed by
+    /// `add_intermediate_aggregation_result`.
+    buckets: Vec<Option<CardinalityCollector>>,
+    accessor_idx: usize,
+    /// The column accessor to access the fast field values.
+    accessor: Arc<dyn ValueSource>,
+    /// The column_type of the field.
+    column_type: ColumnType,
+    /// Set iff `column_type == ColumnType::IpAddr`. Resolved once at
+    /// construction: the raw column values are compact-space codes that must
+    /// be expanded to their u128 ip representation before hashing.
+    compact_space_accessor: Option<Arc<CompactSpaceU64Accessor>>,
+    /// The missing value normalized to the internal u64 representation of the field type.
+    missing_value_for_accessor: Option<u64>,
+}
+
+impl Debug for SegmentNumericCardinalityCollector {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        f.debug_struct("SegmentNumericCardinalityCollector")
+            .field("column_type", &self.column_type)
+            .field(
+                "missing_value_for_accessor",
+                &self.missing_value_for_accessor,
+            )
+            .finish()
+    }
+}
+
+impl SegmentNumericCardinalityCollector {
+    pub fn from_req(
+        accessor_idx: usize,
+        accessor: Arc<dyn ValueSource>,
+        missing_value_for_accessor: Option<u64>,
+    ) -> crate::Result<Self> {
+        let column_type = accessor.column_type();
+        assert_ne!(column_type, ColumnType::Str);
+        let compact_space_accessor = if column_type == ColumnType::IpAddr {
+            let compact_space_accessor = accessor
+                .as_column()
+                .ok_or_else(|| {
+                    TantivyError::AggregationError(
+                        crate::aggregation::AggregationError::InternalError(
+                            "IpAddr cardinality requires a physical column".to_string(),
+                        ),
+                    )
+                })?
+                .values
+                .clone()
+                .downcast_arc::<CompactSpaceU64Accessor>()
+                .map_err(|_| {
+                    TantivyError::AggregationError(
+                        crate::aggregation::AggregationError::InternalError(
+                            "Type mismatch: Could not downcast to CompactSpaceU64Accessor"
+                                .to_string(),
+                        ),
+                    )
+                })?;
+            Some(compact_space_accessor)
+        } else {
+            None
+        };
+        Ok(Self {
+            buckets: Vec::new(),
+            accessor_idx,
+            accessor,
+            column_type,
+            compact_space_accessor,
+            missing_value_for_accessor,
+        })
+    }
+}
+
+impl SegmentAggregationCollector for SegmentNumericCardinalityCollector {
+    fn add_intermediate_aggregation_result(
+        &mut self,
+        agg_data: &AggregationsSegmentCtx,
+        results: &mut IntermediateAggregationResults,
+        bucket_id: BucketId,
+    ) -> crate::Result<()> {
+        self.prepare_max_bucket(bucket_id, agg_data)?;
+        let name = agg_data
+            .get_cardinality_req_data(self.accessor_idx)
+            .name
+            .to_string();
+        // take the bucket in buckets and replace it with a new empty one
+        let Some(cardinality) = self.buckets[bucket_id as usize].take() else {
+            return Err(crate::TantivyError::InternalError(
+                "the same bucket should not be finalized twice.".to_string(),
+            ));
+        };
+        results.push(
+            name,
+            IntermediateAggregationResult::Metric(IntermediateMetricResult::Cardinality(
+                cardinality,
+            )),
+        )?;
+        Ok(())
+    }
+
+    fn collect(
+        &mut self,
+        parent_bucket_id: BucketId,
+        docs: &[crate::DocId],
+        agg_data: &mut AggregationsSegmentCtx,
+    ) -> crate::Result<()> {
+        agg_data.column_block_accessor.fetch_block_with_missing(
+            docs,
+            &*self.accessor,
+            self.missing_value_for_accessor,
+        );
+        let cardinality = self.buckets[parent_bucket_id as usize]
+            .as_mut()
+            .ok_or_else(|| {
+                crate::TantivyError::InternalError(
+                    "collection should not happen after finalization".to_string(),
+                )
+            })?;
+        let col_block_accessor = &agg_data.column_block_accessor;
+        if let Some(compact_space_accessor) = self.compact_space_accessor.as_ref() {
+            for val in col_block_accessor.iter_vals() {
+                let val: u128 = compact_space_accessor.compact_to_u128(val as u32);
+                cardinality.insert(val);
+            }
+        } else {
+            for val in col_block_accessor.iter_vals() {
+                cardinality.insert(val);
+            }
+        }
+        Ok(())
+    }
+
+    fn prepare_max_bucket(
+        &mut self,
+        max_bucket: BucketId,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> crate::Result<()> {
+        if max_bucket as usize >= self.buckets.len() {
+            let column_type = self.column_type;
+            self.buckets.resize_with(max_bucket as usize + 1, || {
+                Some(CardinalityCollector::new(column_type as u8))
+            });
+        }
+        Ok(())
+    }
+
+    fn compute_metric_value(
+        &self,
+        bucket_id: BucketId,
+        sub_agg_name: &str,
+        sub_agg_property: &str,
+        agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        let req_data = &agg_data.get_cardinality_req_data(self.accessor_idx);
+        if req_data.name != sub_agg_name || !sub_agg_property.is_empty() {
+            return None;
+        }
+        let cardinality = self.buckets.get(bucket_id as usize)?.as_ref()?;
+        Some(cardinality.sketch.estimate().trunc())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::net::IpAddr;
+    use std::str::FromStr;
+
+    use crate::aggregation::agg_req::Aggregations;
+    use crate::aggregation::tests::exec_request;
+    use crate::schema::{IntoIpv6Addr, Schema, FAST};
+    use crate::Index;
+
+    #[test]
+    fn cardinality_aggregation_u64() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let id_field = schema_builder.add_u64_field("id", FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        {
+            let mut writer = index.writer_for_tests()?;
+            writer.add_document(doc!(id_field => 1u64))?;
+            writer.add_document(doc!(id_field => 2u64))?;
+            writer.add_document(doc!(id_field => 3u64))?;
+            writer.add_document(doc!())?;
+            writer.commit()?;
+        }
+
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "cardinality": {
+                "cardinality": {
+                    "field": "id",
+                    "missing": 0u64
+                },
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["cardinality"]["value"], 4.0);
+
+        Ok(())
+    }
+
+    #[test]
+    fn cardinality_aggregation_ip_addr() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let field = schema_builder.add_ip_addr_field("ip_field", FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        {
+            let mut writer = index.writer_for_tests()?;
+            // IpV6 loopback
+            writer.add_document(doc!(field=>IpAddr::from_str("::1").unwrap().into_ipv6_addr()))?;
+            writer.add_document(doc!(field=>IpAddr::from_str("::1").unwrap().into_ipv6_addr()))?;
+            // IpV4
+            writer.add_document(
+                doc!(field=>IpAddr::from_str("127.0.0.1").unwrap().into_ipv6_addr()),
+            )?;
+            writer.commit()?;
+        }
+
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "cardinality": {
+                "cardinality": {
+                    "field": "ip_field"
+                },
+            }
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["cardinality"]["value"], 2.0);
+
+        Ok(())
+    }
+}
diff --git a/src/aggregation/metric/cardinality/str_collector.rs b/src/aggregation/metric/cardinality/str_collector.rs
new file mode 100644
index 000000000..b2ec7d218
--- /dev/null
+++ b/src/aggregation/metric/cardinality/str_collector.rs
@@ -0,0 +1,399 @@
+//! Segment collector for `cardinality` over a `ColumnType::Str` column.
+//!
+//! Strings are dictionary encoded, and resolving a term ordinal to its bytes
+//! is the expensive part. So instead of hashing values during collection, the
+//! collector accumulates *term ordinals* per bucket and, at finalization,
+//! builds one shared term_ord -> coupon cache for every bucket at once. The
+//! coupons are then appended into a [`CardinalityCollector`], which is the
+//! same type the numeric collector produces — merging across segments and
+//! across column kinds stays in the parent module.
+
+use std::fmt::Debug;
+use std::io;
+use std::sync::Arc;
+
+use columnar::{ColumnType, Dictionary};
+use datasketches::hll::Coupon;
+use rustc_hash::{FxBuildHasher, FxHashMap, FxHashSet};
+
+use super::term_ord_accumulator::TermOrdAccumulator;
+use super::CardinalityCollector;
+use crate::aggregation::agg_data::AggregationsSegmentCtx;
+use crate::aggregation::intermediate_agg_result::{
+    IntermediateAggregationResult, IntermediateAggregationResults, IntermediateMetricResult,
+};
+use crate::aggregation::segment_agg_result::SegmentAggregationCollector;
+use crate::aggregation::value_source::ValueSource;
+use crate::aggregation::*;
+
+/// A CouponCache is here to cache the mapping term ordinal -> coupon (see above).
+/// The idea is that we do not want to fetch terms associated to several term ordinals,
+/// several times due to the fact that we have several buckets.
+enum CouponCache {
+    Dense {
+        coupon_map: Vec<Coupon>,
+        missing_coupon_opt: Option<Coupon>,
+    },
+    Sparse {
+        coupon_map: FxHashMap<u64, Coupon>,
+        missing_coupon_opt: Option<Coupon>,
+    },
+}
+
+impl CouponCache {
+    fn new(
+        term_ords: Vec<u64>,
+        coupons: Vec<Coupon>,
+        missing_coupon_opt: Option<Coupon>,
+    ) -> CouponCache {
+        let num_terms = term_ords.len();
+        assert_eq!(num_terms, coupons.len());
+        if term_ords.is_empty() {
+            return CouponCache::Dense {
+                coupon_map: Vec::new(),
+                missing_coupon_opt,
+            };
+        }
+        let highest_term_ord = term_ords.last().copied().unwrap_or(0u64);
+        // We prefer the dense implementation, if it is not too wasteful.
+        // There are two cases for which we can use it.
+        // 1- if the data is small.
+        // 2- if the data is not necessarily small, but due to a high occupancy ratio, the RAM usage
+        // is not that much bigger than if we had used a HashSet. (occupancy ratio + extra
+        // metadata ~ x2.25)
+        let should_use_dense =
+            highest_term_ord < 1_000_000u64 || highest_term_ord < num_terms as u64 * 3u64;
+        if should_use_dense {
+            // We don't really care about the value here. We will populate all the values we will
+            // read anyway.
+            let uninitialized_coupon = Coupon::from_value(0);
+            let mut coupon_map: Vec<Coupon> =
+                vec![uninitialized_coupon; highest_term_ord as usize + 1];
+
+            for (term_ord, coupon) in term_ords.into_iter().zip(coupons) {
+                coupon_map[term_ord as usize] = coupon;
+            }
+            CouponCache::Dense {
+                coupon_map,
+                missing_coupon_opt,
+            }
+        } else {
+            let coupon_map: FxHashMap<u64, Coupon> = term_ords.into_iter().zip(coupons).collect();
+            CouponCache::Sparse {
+                coupon_map,
+                missing_coupon_opt,
+            }
+        }
+    }
+}
+
+/// Segment collector for `cardinality` over a `ColumnType::Str` column.
+///
+/// Hidden contract: the column passed at construction must be a str column
+/// whose values are term ordinals of the associated dictionary, and
+/// `max_term_ord_inclusive` must be `accessor.max_value()`. The missing
+/// sentinel `accessor.max_value() + 1` may additionally be inserted when
+/// `missing_value_for_accessor` is set, hence accumulators size for
+/// `max_term_ord_inclusive + 1`.
+pub(crate) struct SegmentStrCardinalityCollector<S: TermOrdAccumulator> {
+    /// Buckets are Some(_) until they get consumed by
+    /// `add_intermediate_aggregation_result`.
+    buckets: Vec<Option<S>>,
+    accessor_idx: usize,
+    /// The column accessor to access the fast field values (term ordinals).
+    accessor: Arc<dyn ValueSource>,
+    /// The missing value normalized to the internal u64 representation of the field type.
+    missing_value_for_accessor: Option<u64>,
+    /// Lazily built at finalization time, shared by every bucket.
+    coupon_cache: Option<CouponCache>,
+    /// Largest term_ord that may be inserted into a bucket, i.e.
+    /// `accessor.max_value()`.
+    max_term_ord_inclusive: u64,
+}
+
+impl<S: TermOrdAccumulator> Debug for SegmentStrCardinalityCollector<S> {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        f.debug_struct("SegmentStrCardinalityCollector")
+            .field("num_buckets", &self.buckets.len())
+            .field(
+                "missing_value_for_accessor",
+                &self.missing_value_for_accessor,
+            )
+            .finish()
+    }
+}
+
+/// Builds a coupon cache from the given buckets, dictionary, and optional missing value.
+/// Returns a mapping from term_ord to the hash (coupon) of the associated term.
+fn build_coupon_cache<S: TermOrdAccumulator>(
+    buckets: &[Option<S>],
+    dictionary: &Dictionary,
+    missing_value_opt: Option<&Key>,
+) -> io::Result<CouponCache> {
+    // Pass 1 computes the capacity hint, pass 2 inserts.
+    let mut max_bucket_len = 0usize;
+    for bucket in buckets.iter().flatten() {
+        max_bucket_len = max_bucket_len.max(bucket.len());
+    }
+    let mut term_ords_set = FxHashSet::with_capacity_and_hasher(max_bucket_len * 2, FxBuildHasher);
+    for bucket in buckets.iter().flatten() {
+        term_ords_set.extend(bucket.iter_ords());
+    }
+    let mut term_ords: Vec<u64> = term_ords_set.into_iter().collect();
+    term_ords.sort_unstable();
+
+    term_ords.pop_if(|highest_term_ord| *highest_term_ord >= dictionary.num_terms() as u64);
+
+    let mut coupons: Vec<Coupon> = Vec::with_capacity(term_ords.len());
+    let all_term_ords_found: bool =
+        dictionary.sorted_ords_to_term_cb(&term_ords, |term_bytes| {
+            let coupon: Coupon = Coupon::from_value(term_bytes);
+            coupons.push(coupon);
+        })?;
+    assert!(all_term_ords_found);
+
+    // Regardless of whether or not there is effectively a missing value in one of the buckets,
+    // we populate the cache with the missing key too (if any).
+    let missing_coupon_opt: Option<Coupon> = missing_value_opt.map(|missing_key| {
+        if let Key::Str(missing_value_str) = missing_key {
+            Coupon::from_value(missing_value_str.as_bytes())
+        } else {
+            // See https://github.com/quickwit-oss/tantivy/issues/2891
+            // A missing key with a type different from Str will not work as intended
+            // for the moment.
+            //
+            // Right now this is just a partial workaround.
+            Coupon::from_value("__tantivy_missing_non_str__".as_bytes())
+        }
+    });
+    Ok(CouponCache::new(term_ords, coupons, missing_coupon_opt))
+}
+
+fn append_to_sketch(
+    term_ords: &impl TermOrdAccumulator,
+    coupon_cache: &CouponCache,
+    sketch: &mut CardinalityCollector,
+) {
+    match coupon_cache {
+        CouponCache::Dense {
+            coupon_map,
+            missing_coupon_opt,
+        } => {
+            if let Some(missing_coupon) = missing_coupon_opt {
+                for term_ord in term_ords.iter_ords() {
+                    let coupon: Coupon = coupon_map
+                        .get(term_ord as usize)
+                        .copied()
+                        .unwrap_or(*missing_coupon);
+                    sketch.insert_coupon(coupon);
+                }
+            } else {
+                for term_ord in term_ords.iter_ords() {
+                    if let Some(coupon) = coupon_map.get(term_ord as usize).copied() {
+                        sketch.insert_coupon(coupon);
+                    }
+                }
+            }
+        }
+        CouponCache::Sparse {
+            coupon_map,
+            missing_coupon_opt,
+        } => {
+            for term_ord in term_ords.iter_ords() {
+                if let Some(coupon) = coupon_map.get(&term_ord).copied().or(*missing_coupon_opt) {
+                    sketch.insert_coupon(coupon);
+                }
+            }
+        }
+    }
+}
+
+impl<S: TermOrdAccumulator> SegmentStrCardinalityCollector<S> {
+    pub fn from_req(
+        accessor_idx: usize,
+        accessor: Arc<dyn ValueSource>,
+        missing_value_for_accessor: Option<u64>,
+        max_term_ord_inclusive: u64,
+    ) -> Self {
+        Self {
+            buckets: Vec::new(),
+            accessor_idx,
+            accessor,
+            missing_value_for_accessor,
+            coupon_cache: None,
+            max_term_ord_inclusive,
+        }
+    }
+}
+
+impl<S: TermOrdAccumulator + 'static> SegmentAggregationCollector
+    for SegmentStrCardinalityCollector<S>
+{
+    fn add_intermediate_aggregation_result(
+        &mut self,
+        agg_data: &AggregationsSegmentCtx,
+        results: &mut IntermediateAggregationResults,
+        bucket_id: BucketId,
+    ) -> crate::Result<()> {
+        self.prepare_max_bucket(bucket_id, agg_data)?;
+        let req_data = &agg_data.get_cardinality_req_data(self.accessor_idx);
+        let Some(str_dict_column) = &req_data.str_dict_column else {
+            return Err(crate::TantivyError::InternalError(
+                "a str cardinality collector requires a str dictionary column".to_string(),
+            ));
+        };
+        // Strings are dictionary encoded. Fetching the terms associated to strings
+        // is expensive. For this reason, we do that once for all buckets and cache the results
+        // here.
+        //
+        // The cache maps a term_ord to the hash of the associated term. The missing value
+        // sentinel will be associated to the hash of the missing value if any.
+        if self.coupon_cache.is_none() {
+            self.coupon_cache = Some(build_coupon_cache(
+                &self.buckets,
+                str_dict_column.dictionary(),
+                req_data.req.missing.as_ref(),
+            )?);
+        }
+        let name = req_data.name.to_string();
+        // take the bucket in buckets and replace it with a new empty one
+        let Some(term_ords) = self.buckets[bucket_id as usize].take() else {
+            return Err(crate::TantivyError::InternalError(
+                "the same bucket should not be finalized twice.".to_string(),
+            ));
+        };
+        let mut cardinality = CardinalityCollector::new(ColumnType::Str as u8);
+        if let Some(coupon_cache) = self.coupon_cache.as_ref() {
+            append_to_sketch(&term_ords, coupon_cache, &mut cardinality);
+        }
+        results.push(
+            name,
+            IntermediateAggregationResult::Metric(IntermediateMetricResult::Cardinality(
+                cardinality,
+            )),
+        )?;
+
+        Ok(())
+    }
+
+    fn collect(
+        &mut self,
+        parent_bucket_id: BucketId,
+        docs: &[crate::DocId],
+        agg_data: &mut AggregationsSegmentCtx,
+    ) -> crate::Result<()> {
+        agg_data.column_block_accessor.fetch_block_with_missing(
+            docs,
+            &*self.accessor,
+            self.missing_value_for_accessor,
+        );
+        let Some(term_ords) = self.buckets[parent_bucket_id as usize].as_mut() else {
+            return Err(crate::TantivyError::InternalError(
+                "collection should not happen after finalization".to_string(),
+            ));
+        };
+        // Promotion check runs on the pre-block state: the first call
+        // sees an empty set (no-op), and the last block of inserts
+        // doesn't trigger a promotion of a set we won't grow further.
+        // The trait dispatches once per block (via `extend_from_iter`)
+        // for adaptive variants and inlines to a tight loop for the
+        // BitSet path.
+        term_ords.maybe_compact();
+        term_ords.extend_from_iter(agg_data.column_block_accessor.iter_vals());
+        Ok(())
+    }
+
+    fn prepare_max_bucket(
+        &mut self,
+        max_bucket: BucketId,
+        _agg_data: &AggregationsSegmentCtx,
+    ) -> crate::Result<()> {
+        if max_bucket as usize >= self.buckets.len() {
+            let max_term_ord_inclusive = self.max_term_ord_inclusive;
+            self.buckets.resize_with(max_bucket as usize + 1, || {
+                Some(S::new(max_term_ord_inclusive))
+            });
+        }
+        Ok(())
+    }
+
+    fn compute_metric_value(
+        &self,
+        bucket_id: BucketId,
+        sub_agg_name: &str,
+        sub_agg_property: &str,
+        agg_data: &AggregationsSegmentCtx,
+    ) -> Option<f64> {
+        let req_data = &agg_data.get_cardinality_req_data(self.accessor_idx);
+        if req_data.name != sub_agg_name || !sub_agg_property.is_empty() {
+            return None;
+        }
+        // The sketch isn't built until finalization; the term_ord set's len is
+        // the exact distinct count.
+        let term_ords = self.buckets.get(bucket_id as usize)?.as_ref()?;
+        Some(term_ords.len() as f64)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::aggregation::agg_req::Aggregations;
+    use crate::aggregation::tests::{exec_request, get_test_index_from_terms};
+    use crate::schema::{Schema, FAST, STRING};
+    use crate::Index;
+
+    /// Build a single-segment string-cardinality index with 32 unique terms.
+    /// `column.max_value() = 31` is well below `BITSET_MAX_TERM_ORD`,
+    /// so the bucket exercises the `BitSet` path end to end.
+    #[test]
+    fn cardinality_aggregation_test_str_bitset() -> crate::Result<()> {
+        let terms: Vec<String> = (0..32).map(|i| format!("term_{i}")).collect();
+        let term_refs: Vec<Vec<&str>> = terms.iter().map(|t| vec![t.as_str()]).collect::<Vec<_>>();
+        // single segment so we have a single dictionary of 32 terms.
+        let index = get_test_index_from_terms(true, &term_refs)?;
+
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "cardinality": {
+                "cardinality": { "field": "string_id" }
+            },
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index)?;
+        assert_eq!(res["cardinality"]["value"], 32.0);
+        Ok(())
+    }
+
+    /// `BitSet` path with a `missing` parameter: the column-level missing
+    /// sentinel (`column.max_value() + 1`) flows into the bitset, the
+    /// dict lookup filter at finalization drops it, and the missing
+    /// coupon is applied separately.
+    #[test]
+    fn cardinality_aggregation_test_str_bitset_with_missing() {
+        let mut schema_builder = Schema::builder();
+        let name_field = schema_builder.add_text_field("name", STRING | FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut writer = index.writer_for_tests().unwrap();
+        for i in 0..16 {
+            let term = format!("t{i:02}");
+            writer.add_document(doc!(name_field => term)).unwrap();
+        }
+        // One empty doc, exercising the missing sentinel.
+        writer.add_document(doc!()).unwrap();
+        writer.commit().unwrap();
+
+        let agg_req: Aggregations = serde_json::from_value(json!({
+            "cardinality": {
+                "cardinality": {
+                    "field": "name",
+                    "missing": "MISSING_SENTINEL_KEY",
+                }
+            },
+        }))
+        .unwrap();
+
+        let res = exec_request(agg_req, &index).unwrap();
+        // 16 distinct real terms + 1 distinct "missing" value = 17.
+        assert_eq!(res["cardinality"]["value"], 17.0);
+    }
+}
diff --git a/src/aggregation/metric/cardinality/term_ord_accumulator.rs b/src/aggregation/metric/cardinality/term_ord_accumulator.rs
new file mode 100644
index 000000000..f35d6148c
--- /dev/null
+++ b/src/aggregation/metric/cardinality/term_ord_accumulator.rs
@@ -0,0 +1,343 @@
+//! Per-bucket term-ordinal accumulators used by the str cardinality
+//! collector.
+//!
+//! A str cardinality bucket needs a *set of term ordinals*, and the right
+//! representation depends on how large the segment's dictionary slice is:
+//!
+//!   * [`BitSet`] (from `common`): used when `column.max_value()` is small (<
+//!     [`BITSET_MAX_TERM_ORD`]). Pre-allocated, no promotion machinery.
+//!   * [`TermOrdSet`]: adaptive. Starts as an `FxHashSet` (cheap when few ords are seen) and
+//!     promotes to a [`PagedBitset`] once occupancy crosses the density threshold.
+//!
+//! Both are exposed through the [`TermOrdAccumulator`] trait so that
+//! `SegmentStrCardinalityCollector` can be generic over the choice and the hot
+//! `collect()` loop monomorphizes to a direct call (no enum dispatch per
+//! insert).
+
+use common::{BitSet, TinySet};
+use rustc_hash::FxHashSet;
+
+/// Promote FxHashSet<u64> -> PagedBitset at ~3% density (`len * 32 >
+/// dict_num_terms`). Past this point the bitset (~`dict_num_terms / 7.5`
+/// bytes) is smaller than the hashset (~10 B/entry minimum) and avoids
+/// the per-insert hash.
+const PROMOTION_RATIO: u64 = 32;
+
+// =================================================================
+// PagedBitset: a sparse bitset indexed by term_ord.
+//
+// Used as the dense alternative to FxHashSet<u64> once a string
+// cardinality bucket has accumulated enough unique term ordinals.
+// Memory is bounded to (touched pages) * (page bytes), not
+// (max_term_ord / 8).
+//
+// Page geometry mirrors `PagedTermMap` in `term_agg.rs`: 1024 ords
+// per page, lazy `Vec<Option<Box<Page>>>` directory.
+// =================================================================
+const BITSET_PAGE_SHIFT: u32 = 10;
+const BITSET_PAGE_BITS: u64 = 1u64 << BITSET_PAGE_SHIFT; // 1024
+const BITSET_PAGE_MASK: u64 = BITSET_PAGE_BITS - 1;
+const BITSET_WORDS_PER_PAGE: usize = (BITSET_PAGE_BITS / 64) as usize; // 16
+
+#[derive(Clone)]
+struct PagedBitsetPage {
+    words: [TinySet; BITSET_WORDS_PER_PAGE],
+}
+
+impl PagedBitsetPage {
+    fn new() -> Self {
+        Self {
+            words: [TinySet::empty(); BITSET_WORDS_PER_PAGE],
+        }
+    }
+}
+
+pub(crate) struct PagedBitset {
+    pages: Vec<Option<Box<PagedBitsetPage>>>,
+    /// Cached number of set bits, maintained on insert.
+    count: u64,
+}
+
+impl PagedBitset {
+    /// Allocates a directory big enough to hold ords up to and including
+    /// `max_term_ord`. Pages are allocated lazily on first set.
+    fn with_max_term_ord(max_term_ord: u64) -> Self {
+        let max_page_idx = (max_term_ord >> BITSET_PAGE_SHIFT) as usize;
+        let num_pages = max_page_idx + 1;
+        Self {
+            pages: vec![None; num_pages],
+            count: 0,
+        }
+    }
+
+    #[inline]
+    fn insert(&mut self, term_ord: u64) {
+        let page_idx = (term_ord >> BITSET_PAGE_SHIFT) as usize;
+        let intra = term_ord & BITSET_PAGE_MASK;
+        let word_idx = (intra >> 6) as usize;
+        let bit_idx = (intra & 63) as u32;
+
+        let page = match &mut self.pages[page_idx] {
+            Some(p) => p,
+            None => {
+                self.pages[page_idx] = Some(Box::new(PagedBitsetPage::new()));
+                self.pages[page_idx].as_mut().unwrap()
+            }
+        };
+        if page.words[word_idx].insert_mut(bit_idx) {
+            self.count += 1;
+        }
+    }
+
+    /// Number of set bits. O(1).
+    #[inline]
+    fn len(&self) -> u64 {
+        self.count
+    }
+
+    /// Iterate set ords in ascending order.
+    fn iter_sorted(&self) -> impl Iterator<Item = u64> + '_ {
+        self.pages
+            .iter()
+            .enumerate()
+            .filter_map(|(page_idx, page_opt)| page_opt.as_ref().map(|p| (page_idx, p)))
+            .flat_map(|(page_idx, page)| {
+                let page_base_ord = (page_idx as u64) << BITSET_PAGE_SHIFT;
+                page.words
+                    .iter()
+                    .enumerate()
+                    .flat_map(move |(word_idx, &word)| {
+                        let word_base_ord = page_base_ord + (word_idx as u64) * 64;
+                        word.into_iter()
+                            .map(move |bit| word_base_ord + u64::from(bit))
+                    })
+            })
+    }
+}
+
+/// Threshold below which we use `BitSet` instead of `TermOrdSet`.
+///
+/// Both `BitSet` and `FxHashSet<u64>` have the same 32-byte struct, so the comparison is heap only:
+///   * `BitSet` at T=256: 5 `TinySet` words covering 258 bits (with the missing-value sentinel) =
+///     40 bytes.
+///   * `FxHashSet<u64>` after one insert: 4-bucket hashbrown table ≈ 56 bytes
+pub(crate) const BITSET_MAX_TERM_ORD: u64 = 256;
+
+// =================================================================
+// TermOrdAccumulator: per-bucket abstraction over the entries set.
+//
+// Implementations:
+//   - `BitSet` (from `common`): used when `column.max_value()` is small (< BITSET_MAX_TERM_ORD).
+//     Pre-allocated, no promotion.
+//   - `TermOrdSet`: adaptive, starts as FxHashSet and promotes to a paged bitset when occupancy
+//     crosses the density threshold (only if promotion is enabled — typically gated on top-level
+//     aggregation).
+//
+// The trait lets `SegmentStrCardinalityCollector` be generic over the choice
+// so the hot collect() loop monomorphizes to a direct call (no enum
+// dispatch per insert).
+// =================================================================
+pub(crate) trait TermOrdAccumulator: Sized {
+    /// Construct an empty accumulator.
+    /// `max_term_ord_inclusive` is the largest term_ord that may be
+    /// inserted (used to size pre-allocated bitsets and the dense bitset
+    /// on promotion).
+    fn new(max_term_ord_inclusive: u64) -> Self;
+    fn insert(&mut self, term_ord: u64);
+    fn extend_from_iter(&mut self, ords: impl IntoIterator<Item = u64>);
+    /// Hook called once per ingested block. Adaptive impls use this to
+    /// decide on sparse->dense promotion.
+    fn maybe_compact(&mut self) {}
+    fn len(&self) -> usize;
+    fn iter_ords(&self) -> impl Iterator<Item = u64> + '_;
+}
+
+impl TermOrdAccumulator for BitSet {
+    #[inline]
+    fn new(max_term_ord_inclusive: u64) -> Self {
+        // `BitSet::with_max_value(M)` accepts ords in [0, M).
+        // We need ords up to and including `max_term_ord_inclusive`, plus
+        // the missing-value sentinel `column.max_value() + 1`.
+        BitSet::with_max_value((max_term_ord_inclusive + 2) as u32)
+    }
+    #[inline]
+    fn insert(&mut self, term_ord: u64) {
+        BitSet::insert(self, term_ord as u32);
+    }
+    #[inline]
+    fn len(&self) -> usize {
+        BitSet::len(self)
+    }
+    fn iter_ords(&self) -> impl Iterator<Item = u64> + '_ {
+        // `BitSet` itself doesn't expose iteration, but
+        // `BitSet::tinyset(bucket)` does. Walk per-bucket and yield each
+        // set bit. The capacity is `max_value()`; iterating to
+        // `div_ceil(64)` covers every possible ord exactly once.
+        let num_buckets = self.max_value().div_ceil(64);
+        (0..num_buckets).flat_map(move |bucket| {
+            let chunk_base = u64::from(bucket) * 64;
+            self.tinyset(bucket)
+                .into_iter()
+                .map(move |bit| chunk_base + u64::from(bit))
+        })
+    }
+    #[inline(never)] //< required to not have a perf regression
+    fn extend_from_iter(&mut self, ords: impl IntoIterator<Item = u64>) {
+        for ord in ords {
+            <Self as TermOrdAccumulator>::insert(self, ord);
+        }
+    }
+}
+
+// TermOrdSet: adaptive sparse->dense accumulator.
+//
+// Starts as an HashSet (cheap when few ords are seen). When occupancy
+// crosses `len * PROMOTION_RATIO > max_term_ord_inclusive`, drains into
+// a `PagedBitset` and continues dense.
+pub(crate) struct TermOrdSet {
+    inner: TermOrdSetInner,
+    /// Largest term_ord that may be inserted. Used for both sizing the
+    /// dense bitset on promotion and as the promotion-threshold reference.
+    max_term_ord_inclusive: u64,
+}
+
+enum TermOrdSetInner {
+    Sparse(FxHashSet<u64>),
+    Dense(PagedBitset),
+}
+
+impl TermOrdAccumulator for TermOrdSet {
+    fn new(max_term_ord_inclusive: u64) -> Self {
+        Self {
+            inner: TermOrdSetInner::Sparse(FxHashSet::default()),
+            max_term_ord_inclusive,
+        }
+    }
+
+    #[inline]
+    fn insert(&mut self, term_ord: u64) {
+        match &mut self.inner {
+            TermOrdSetInner::Sparse(set) => {
+                set.insert(term_ord);
+            }
+            TermOrdSetInner::Dense(bitset) => bitset.insert(term_ord),
+        }
+    }
+
+    fn extend_from_iter(&mut self, ords: impl IntoIterator<Item = u64>) {
+        match &mut self.inner {
+            TermOrdSetInner::Sparse(set) => {
+                set.extend(ords);
+            }
+            TermOrdSetInner::Dense(bitset) => {
+                for ord in ords {
+                    bitset.insert(ord);
+                }
+            }
+        }
+    }
+
+    fn maybe_compact(&mut self) {
+        let TermOrdSetInner::Sparse(set) = &mut self.inner else {
+            return;
+        };
+        if set.len() as u64 * PROMOTION_RATIO <= self.max_term_ord_inclusive {
+            return;
+        }
+        let mut bitset = PagedBitset::with_max_term_ord(self.max_term_ord_inclusive + 1);
+        let set = std::mem::take(set);
+        for ord in set {
+            bitset.insert(ord);
+        }
+        self.inner = TermOrdSetInner::Dense(bitset);
+    }
+
+    fn len(&self) -> usize {
+        match &self.inner {
+            TermOrdSetInner::Sparse(set) => set.len(),
+            TermOrdSetInner::Dense(bitset) => bitset.len() as usize,
+        }
+    }
+
+    fn iter_ords(&self) -> impl Iterator<Item = u64> + '_ {
+        match &self.inner {
+            TermOrdSetInner::Sparse(set) => itertools::Either::Left(set.iter().copied()),
+            TermOrdSetInner::Dense(bitset) => itertools::Either::Right(bitset.iter_sorted()),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use common::BitSet;
+
+    use super::{PagedBitset, TermOrdAccumulator, TermOrdSet, PROMOTION_RATIO};
+
+    /// Unit-test the PagedBitset itself: cross-page inserts produce sorted
+    /// iteration, len() matches the inserted set, and duplicates are
+    /// idempotent.
+    #[test]
+    fn paged_bitset_basic() {
+        // Span several pages: BITSET_PAGE_BITS = 1024, so ords > 1024 land
+        // on the second page, > 2048 on the third, etc.
+        let ords = [0u64, 1, 63, 64, 1023, 1024, 1025, 4096, 4097, 9999, 10_000];
+        let max_ord = *ords.iter().max().unwrap();
+        let mut bitset = PagedBitset::with_max_term_ord(max_ord);
+        for &ord in &ords {
+            bitset.insert(ord);
+            // Idempotent: inserting again must not increase count.
+            bitset.insert(ord);
+        }
+        assert_eq!(bitset.len(), ords.len() as u64);
+        let collected: Vec<u64> = bitset.iter_sorted().collect();
+        let mut expected: Vec<u64> = ords.to_vec();
+        expected.sort_unstable();
+        assert_eq!(collected, expected);
+    }
+
+    /// Unit-test `TermOrdSet`: starts Sparse, promotes to Dense on
+    /// `maybe_compact` once the density threshold is crossed, and
+    /// `iter_ords()` yields the same set in either state. Ords spanning
+    /// multiple paged-bitset pages exercise the Dense iter ordering.
+    #[test]
+    fn term_ord_set_promotes_on_maybe_compact() {
+        // Pick max so promotion needs few inserts: len * RATIO > max with
+        // RATIO=32 and max=64 trips at len=3 (3*32=96 > 64).
+        let max_term_ord = 64u64;
+        let mut set = <TermOrdSet as TermOrdAccumulator>::new(max_term_ord);
+        // Two inserts: should stay Sparse after maybe_compact (2 * RATIO = 64, not > 64).
+        set.insert(0);
+        set.insert(7);
+        set.maybe_compact();
+        assert_eq!(set.len(), 2);
+
+        // Third insert promotes on next maybe_compact.
+        set.insert(20);
+        assert_eq!(set.len(), 3);
+        // Sanity check: at len=3, 3 * PROMOTION_RATIO = 96 > 64.
+        assert!(3u64 * PROMOTION_RATIO > max_term_ord);
+        set.maybe_compact();
+
+        // Post-promotion: extending continues to work.
+        set.insert(15);
+        set.insert(15); // dup
+        assert_eq!(set.len(), 4);
+
+        let mut collected: Vec<u64> = set.iter_ords().collect();
+        collected.sort_unstable();
+        assert_eq!(collected, vec![0, 7, 15, 20]);
+    }
+
+    /// Unit-test the `BitSet` impl of `TermOrdAccumulator`: insert,
+    /// dedup, and iter_ords order.
+    #[test]
+    fn bitset_accumulator_basic() {
+        let mut set = <BitSet as TermOrdAccumulator>::new(255);
+        for ord in [0u64, 1, 63, 64, 65, 128, 200, 200, 0] {
+            <BitSet as TermOrdAccumulator>::insert(&mut set, ord);
+        }
+        assert_eq!(<BitSet as TermOrdAccumulator>::len(&set), 7);
+        let collected: Vec<u64> = set.iter_ords().collect();
+        assert_eq!(collected, vec![0, 1, 63, 64, 65, 128, 200]);
+    }
+}
diff --git a/src/aggregation/metric/extended_stats.rs b/src/aggregation/metric/extended_stats.rs
index 1e625d5de..3747a0921 100644
--- a/src/aggregation/metric/extended_stats.rs
+++ b/src/aggregation/metric/extended_stats.rs
@@ -1,5 +1,6 @@
 use std::fmt::Debug;
 use std::mem;
+use std::sync::Arc;
 
 use serde::{Deserialize, Serialize};
 
@@ -321,8 +322,7 @@ impl IntermediateExtendedStats {
 pub(crate) struct SegmentExtendedStatsCollector {
     name: String,
     missing: Option<u64>,
-    field_type: ColumnType,
-    accessor: columnar::Column<u64>,
+    accessor: Arc<dyn ValueSource>,
     buckets: Vec<IntermediateExtendedStats>,
     sigma: Option<f64>,
 }
@@ -331,10 +331,9 @@ impl SegmentExtendedStatsCollector {
     pub fn from_req(req: &MetricAggReqData, sigma: Option<f64>) -> Self {
         let missing = req
             .missing
-            .and_then(|val| f64_to_fastfield_u64(val, &req.field_type));
+            .and_then(|val| f64_to_fastfield_u64(val, &req.accessor.column_type()));
         Self {
             name: req.name.clone(),
-            field_type: req.field_type,
             accessor: req.accessor.clone(),
             missing,
             buckets: vec![IntermediateExtendedStats::with_sigma(sigma); 16],
@@ -373,11 +372,14 @@ impl SegmentAggregationCollector for SegmentExtendedStatsCollector {
     ) -> crate::Result<()> {
         let mut extended_stats = self.buckets[parent_bucket_id as usize].clone();
 
-        agg_data
-            .column_block_accessor
-            .fetch_block_with_missing(docs, &self.accessor, self.missing);
+        agg_data.column_block_accessor.fetch_block_with_missing(
+            docs,
+            &*self.accessor,
+            self.missing,
+        );
+        let field_type = self.accessor.column_type();
         for val in agg_data.column_block_accessor.iter_vals() {
-            let val1 = f64_from_fastfield_u64(val, self.field_type);
+            let val1 = f64_from_fastfield_u64(val, field_type);
             extended_stats.collect(val1);
         }
 
diff --git a/src/aggregation/metric/mod.rs b/src/aggregation/metric/mod.rs
index 05ff861f2..16197a7a2 100644
--- a/src/aggregation/metric/mod.rs
+++ b/src/aggregation/metric/mod.rs
@@ -28,10 +28,10 @@ mod sum;
 mod top_hits;
 
 use std::collections::HashMap;
+use std::sync::Arc;
 
 pub use average::*;
 pub use cardinality::*;
-use columnar::{Column, ColumnType};
 pub use count::*;
 pub use extended_stats::*;
 pub use max::*;
@@ -43,26 +43,25 @@ pub use stats::*;
 pub use sum::*;
 pub use top_hits::*;
 
+use crate::aggregation::value_source::ValueSource;
 use crate::schema::OwnedValue;
 
 /// Contains all information required by metric aggregations like avg, min, max, sum, stats,
 /// extended_stats, count, percentiles.
 #[repr(C)]
-pub struct MetricAggReqData {
+pub(crate) struct MetricAggReqData {
     /// True if the field is of number or date type.
-    pub is_number_or_date_type: bool,
-    /// The type of the field.
-    pub field_type: ColumnType,
+    pub(crate) is_number_or_date_type: bool,
     /// The missing value normalized to the internal u64 representation of the field type.
-    pub missing_u64: Option<u64>,
+    pub(crate) missing_u64: Option<u64>,
     /// The column accessor to access the fast field values.
-    pub accessor: Column<u64>,
+    pub(crate) accessor: Arc<dyn ValueSource>,
     /// Used when converting to intermediate result
-    pub collecting_for: StatsType,
+    pub(crate) collecting_for: StatsType,
     /// The missing value
-    pub missing: Option<f64>,
+    pub(crate) missing: Option<f64>,
     /// The name of the aggregation.
-    pub name: String,
+    pub(crate) name: String,
 }
 
 impl MetricAggReqData {
diff --git a/src/aggregation/metric/percentiles.rs b/src/aggregation/metric/percentiles.rs
index 9b9c878a6..e2817617e 100644
--- a/src/aggregation/metric/percentiles.rs
+++ b/src/aggregation/metric/percentiles.rs
@@ -1,4 +1,5 @@
 use std::fmt::Debug;
+use std::sync::Arc;
 
 use serde::{Deserialize, Serialize};
 
@@ -134,12 +135,10 @@ impl PercentilesAggregationReq {
 pub(crate) struct SegmentPercentilesCollector {
     pub(crate) buckets: Vec<PercentilesCollector>,
     pub(crate) accessor_idx: usize,
-    /// The type of the field.
-    pub field_type: ColumnType,
     /// The missing value normalized to the internal u64 representation of the field type.
     pub missing_u64: Option<u64>,
     /// The column accessor to access the fast field values.
-    pub accessor: Column<u64>,
+    pub(crate) accessor: Arc<dyn ValueSource>,
 }
 
 #[derive(Clone, Serialize, Deserialize)]
@@ -250,14 +249,12 @@ impl PercentilesCollector {
 
 impl SegmentPercentilesCollector {
     pub fn from_req_and_validate(
-        field_type: ColumnType,
         missing_u64: Option<u64>,
-        accessor: Column<u64>,
+        accessor: Arc<dyn ValueSource>,
         accessor_idx: usize,
     ) -> Self {
         Self {
             buckets: Vec::with_capacity(64),
-            field_type,
             missing_u64,
             accessor,
             accessor_idx,
@@ -299,12 +296,13 @@ impl SegmentAggregationCollector for SegmentPercentilesCollector {
         let percentiles = &mut self.buckets[parent_bucket_id as usize];
         agg_data.column_block_accessor.fetch_block_with_missing(
             docs,
-            &self.accessor,
+            &*self.accessor,
             self.missing_u64,
         );
 
+        let field_type = self.accessor.column_type();
         for val in agg_data.column_block_accessor.iter_vals() {
-            let val1 = f64_from_fastfield_u64(val, self.field_type);
+            let val1 = f64_from_fastfield_u64(val, field_type);
             percentiles.collect(val1);
         }
 
diff --git a/src/aggregation/metric/stats.rs b/src/aggregation/metric/stats.rs
index 06a39f6d1..30ec18e33 100644
--- a/src/aggregation/metric/stats.rs
+++ b/src/aggregation/metric/stats.rs
@@ -1,4 +1,5 @@
 use std::fmt::Debug;
+use std::sync::Arc;
 
 use columnar::{Column, ColumnType};
 use serde::{Deserialize, Serialize};
@@ -205,6 +206,10 @@ fn create_collector<const TYPE_ID: u8>(
         collecting_for: req.collecting_for,
         is_number_or_date_type: req.is_number_or_date_type,
         missing_u64: req.missing_u64,
+        column_opt: req
+            .accessor
+            .as_column()
+            .map(|column| (column.clone(), req.accessor.column_type())),
         accessor: req.accessor.clone(),
         buckets: vec![IntermediateStats::default()],
     })
@@ -214,7 +219,7 @@ fn create_collector<const TYPE_ID: u8>(
 pub(crate) fn build_segment_stats_collector(
     req: &MetricAggReqData,
 ) -> crate::Result<Box<dyn SegmentAggregationCollector>> {
-    match req.field_type {
+    match req.accessor.column_type() {
         ColumnType::I64 => Ok(create_collector::<{ ColumnType::I64 as u8 }>(req)),
         ColumnType::U64 => Ok(create_collector::<{ ColumnType::U64 as u8 }>(req)),
         ColumnType::F64 => Ok(create_collector::<{ ColumnType::F64 as u8 }>(req)),
@@ -230,7 +235,13 @@ pub(crate) fn build_segment_stats_collector(
 #[derive(Clone, Debug)]
 pub(crate) struct SegmentStatsCollector<const COLUMN_TYPE_ID: u8> {
     pub(crate) missing_u64: Option<u64>,
-    pub(crate) accessor: Column<u64>,
+    pub(crate) accessor: Arc<dyn ValueSource>,
+    /// The physical column backing `accessor`, if any, resolved once at construction.
+    ///
+    /// `collect` is called once per bucket, often with a single doc (e.g. under a
+    /// high-cardinality terms agg), so a per-call virtual dispatch through `accessor` is
+    /// measurable.
+    pub(crate) column_opt: Option<(Column<u64>, ColumnType)>,
     pub(crate) is_number_or_date_type: bool,
     pub(crate) buckets: Vec<IntermediateStats>,
     pub(crate) name: String,
@@ -290,21 +301,31 @@ impl<const COLUMN_TYPE_ID: u8> SegmentAggregationCollector
         // skips the block accessor's buffers entirely.
         // Only valid without a missing value: `values_for_doc` yields nothing for a doc without a
         // value, so the substitute would be silently dropped.
+        // Also only valid over a materialized column: `values_for_doc` is per-document random
+        // access, which a computed source cannot offer. Those fall through to the block path
+        // below, which is semantically identical.
         // TODO: remove once we fetch all values for all bucket ids in one go
-        if docs.len() == 1 && self.missing_u64.is_none() {
-            collect_stats::<COLUMN_TYPE_ID>(
-                &mut self.buckets[parent_bucket_id as usize],
-                self.accessor.values_for_doc(docs[0]),
-                self.is_number_or_date_type,
-            )?;
-
-            return Ok(());
+        if let Some(column_source) = &self.column_opt {
+            if docs.len() == 1 && self.missing_u64.is_none() {
+                collect_stats::<COLUMN_TYPE_ID>(
+                    &mut self.buckets[parent_bucket_id as usize],
+                    column_source.0.values_for_doc(docs[0]),
+                    self.is_number_or_date_type,
+                )?;
+                return Ok(());
+            }
+            agg_data.column_block_accessor.fetch_block_with_missing(
+                docs,
+                column_source,
+                self.missing_u64,
+            );
+        } else {
+            agg_data.column_block_accessor.fetch_block_with_missing(
+                docs,
+                &*self.accessor,
+                self.missing_u64,
+            );
         }
-        agg_data.column_block_accessor.fetch_block_with_missing(
-            docs,
-            &self.accessor,
-            self.missing_u64,
-        );
         collect_stats::<COLUMN_TYPE_ID>(
             &mut self.buckets[parent_bucket_id as usize],
             agg_data.column_block_accessor.iter_vals(),
diff --git a/src/aggregation/metric/top_hits.rs b/src/aggregation/metric/top_hits.rs
index 77e2856a4..4cf1cf4eb 100644
--- a/src/aggregation/metric/top_hits.rs
+++ b/src/aggregation/metric/top_hits.rs
@@ -24,17 +24,17 @@ use crate::{DocAddress, DocId, SegmentOrdinal};
 /// Contains all information required by the TopHitsSegmentCollector to perform the
 /// top_hits aggregation on a segment.
 #[derive(Default)]
-pub struct TopHitsAggReqData {
+pub(crate) struct TopHitsAggReqData {
     /// The accessors to access the fast field values.
-    pub accessors: Vec<(Column<u64>, ColumnType)>,
+    pub(crate) accessors: Vec<(Column<u64>, ColumnType)>,
     /// The accessors to access the fast field values for retrieving document fields.
-    pub value_accessors: HashMap<String, Vec<DynamicColumn>>,
+    pub(crate) value_accessors: HashMap<String, Vec<DynamicColumn>>,
     /// The ordinal of the segment this request data is for.
-    pub segment_ordinal: SegmentOrdinal,
+    pub(crate) segment_ordinal: SegmentOrdinal,
     /// The name of the aggregation.
-    pub name: String,
+    pub(crate) name: String,
     /// The top_hits aggregation request.
-    pub req: TopHitsAggregationReq,
+    pub(crate) req: TopHitsAggregationReq,
 }
 
 impl TopHitsAggReqData {
diff --git a/src/aggregation/mod.rs b/src/aggregation/mod.rs
index 39da6f45e..81b03551c 100644
--- a/src/aggregation/mod.rs
+++ b/src/aggregation/mod.rs
@@ -132,7 +132,6 @@ mod agg_data;
 mod agg_limits;
 pub mod agg_req;
 pub mod agg_result;
-mod block_accessor;
 pub mod bucket;
 pub(crate) mod buffered_sub_aggs;
 mod collector;
@@ -142,10 +141,13 @@ pub mod intermediate_agg_result;
 pub mod metric;
 
 mod segment_agg_result;
+mod value_source;
 use std::cmp::Ordering;
 use std::fmt::Display;
+use std::sync::Arc;
 
-pub(crate) use block_accessor::ColumnBlockAccessor;
+pub(crate) use value_source::ColumnBlockAccessor;
+pub use value_source::{ValueSource, ValueSourceProvider, ValueSourceRegistry};
 
 #[cfg(test)]
 mod agg_tests;
@@ -184,18 +186,31 @@ pub type BucketId = u32;
 /// This struct holds shared resources needed during aggregation execution:
 /// - `limits`: Memory and bucket limits for the aggregation
 /// - `tokenizers`: TokenizerManager for parsing query strings in filter aggregations
+/// - `value_sources`: Named computed columns that aggregations may read instead of a fast field
 #[derive(Clone, Default)]
 pub struct AggContextParams {
     /// Aggregation limits (memory and bucket count)
     pub limits: AggregationLimitsGuard,
     /// Tokenizer manager for query string parsing
     pub tokenizers: TokenizerManager,
+    /// Computed columns registered by name, resolved in preference to a fast field.
+    pub value_sources: Arc<ValueSourceRegistry>,
 }
 
 impl AggContextParams {
     /// Create new aggregation context parameters
     pub fn new(limits: AggregationLimitsGuard, tokenizers: TokenizerManager) -> Self {
-        Self { limits, tokenizers }
+        Self {
+            limits,
+            tokenizers,
+            value_sources: Arc::new(ValueSourceRegistry::default()),
+        }
+    }
+
+    /// Attaches named computed columns, which aggregation requests address as field names.
+    pub fn with_value_sources(mut self, value_sources: Arc<ValueSourceRegistry>) -> Self {
+        self.value_sources = value_sources;
+        self
     }
 }
 
diff --git a/src/aggregation/block_accessor.rs b/src/aggregation/value_source/block_accessor.rs
similarity index 74%
rename from src/aggregation/block_accessor.rs
rename to src/aggregation/value_source/block_accessor.rs
index 2eb592f1b..ccc6b9951 100644
--- a/src/aggregation/block_accessor.rs
+++ b/src/aggregation/value_source/block_accessor.rs
@@ -1,26 +1,11 @@
 use std::cmp::Ordering;
 
-use columnar::{Cardinality, Column, RowId};
+use columnar::{Cardinality, ColumnValues, RowId};
 
+use crate::aggregation::value_source::ValueSource;
 use crate::DocId;
 
-/// A source of values for a block of documents.
-///
-/// Implementations replace the contents of `values` and, for non-full sources, `docids`. The
-/// returned cardinality describes how the two buffers are aligned. Full sources must return one
-/// value per input document in the same order. Optional and multivalued sources must populate
-/// `docids` with one document id per value.
-pub(crate) trait BlockValueSource {
-    fn load_block(
-        &self,
-        docs: &[DocId],
-        values: &mut Vec<u64>,
-        docids: &mut Vec<DocId>,
-        row_ids: &mut Vec<RowId>,
-    ) -> Cardinality;
-}
-
-/// Buffers the values associated with a block of documents loaded from a [`BlockValueSource`].
+/// Buffers the values associated with a block of documents loaded from a [`ValueSource`].
 ///
 /// Regardless of their original types, values are loaded in their `u64` representation using the
 /// associated monotonic mapping.
@@ -29,6 +14,7 @@ pub(crate) struct ColumnBlockAccessor {
     /// Values loaded for the latest document block, in monotonic `u64` representation.
     val_cache: Vec<u64>,
     /// Document ID corresponding to each value in `val_cache` for a non-full source.
+    /// For full sources, this is likely to be empty.
     ///
     /// A document can occur more than once for a multivalued source. For a full source this buffer
     /// is ignored because `val_cache` is aligned directly with the requested document block.
@@ -37,63 +23,56 @@ pub(crate) struct ColumnBlockAccessor {
     missing_docids_cache: Vec<DocId>,
     /// Scratch buffer available to sources for translating document IDs into value row IDs.
     row_id_cache: Vec<RowId>,
-    /// Cardinality reported by the source that loaded the latest block.
-    /// For the moment this is reporting the cardinality of the full column, not
-    /// something specific to the block.
+    /// Cardinality here is describes the relationship with the loaded doc_id_cache and val_cache.
+    ///
+    /// Cheaply hints the cardinality of the given block.
+    ///
+    /// It is to be read as a "lower-bound" hint.
+    /// For instance, a block with one value per doc could have a cardinality property
+    /// set to full, optional or multivalued (both are technically true). For physical column
+    /// for instance, we just set cardinality to the column cardinality (although individual
+    /// blocks could have a stricter cardinality).
+    ///
+    /// See also [`Self::has_one_value_per_doc`] if you need a stricter notion of
+    /// cardinality.
     cardinality: Cardinality,
 }
 
-impl BlockValueSource for Column<u64> {
-    #[inline]
-    fn load_block(
-        &self,
-        docs: &[DocId],
-        values: &mut Vec<u64>,
-        docids: &mut Vec<DocId>,
-        row_ids: &mut Vec<RowId>,
-    ) -> Cardinality {
-        let cardinality = self.index.get_cardinality();
-        if cardinality.is_full() {
-            load_full_column_values(docs, self, values);
-        } else {
-            docids.clear();
-            row_ids.clear();
-            self.row_ids_for_docs(docs, docids, row_ids);
-            values.resize(row_ids.len(), 0u64);
-            self.values.get_vals(row_ids, values);
-        }
-        cardinality
-    }
-}
-
 impl ColumnBlockAccessor {
     #[inline]
-    pub(crate) fn fetch_block(&mut self, docs: &[DocId], source: &impl BlockValueSource) {
+    pub(crate) fn fetch_block<S: ValueSource + ?Sized>(&mut self, docs: &[DocId], source: &S) {
         self.cardinality = source.load_block(
             docs,
             &mut self.val_cache,
             &mut self.docid_cache,
             &mut self.row_id_cache,
         );
+        debug_assert!(
+            !self.cardinality.is_full() || self.val_cache.len() == docs.len(),
+            "a Full source must return exactly one value per input doc"
+        );
     }
 
-    /// Fetches a physical column known to be full without querying its cardinality.
+    /// Fetches a block from a column known to be full (hence we pass the ColumnValue Object
+    /// directly).
     ///
-    /// This direct-column-only entry point is reserved for specialized collectors whose
-    /// construction already proved the column is full.
+    /// docs needs to be strictly increasing.
     #[inline]
-    pub(crate) fn fetch_full_column_block(&mut self, docs: &[DocId], accessor: &Column<u64>) {
-        debug_assert!(accessor.index.get_cardinality().is_full());
-        load_full_column_values(docs, accessor, &mut self.val_cache);
+    pub(crate) fn fetch_full_column_block(
+        &mut self,
+        docs: &[DocId],
+        column_values: &dyn ColumnValues<u64>,
+    ) {
+        super::load_full_column_values(docs, column_values, &mut self.val_cache);
         self.cardinality = Cardinality::Full;
     }
 
     /// Fetches a block and appends `missing_opt` for documents without a value.
     #[inline]
-    pub(crate) fn fetch_block_with_missing(
+    pub(crate) fn fetch_block_with_missing<S: ValueSource + ?Sized>(
         &mut self,
         docs: &[DocId],
-        source: &impl BlockValueSource,
+        source: &S,
         missing_opt: Option<u64>,
     ) {
         self.fetch_block_with_missing_ordered(docs, source, missing_opt, false)
@@ -103,10 +82,10 @@ impl ColumnBlockAccessor {
     /// true, the missing entries are inserted in document order instead of appended as a second
     /// run.
     #[inline]
-    pub(crate) fn fetch_block_with_missing_ordered(
+    pub(crate) fn fetch_block_with_missing_ordered<S: ValueSource + ?Sized>(
         &mut self,
         docs: &[DocId],
-        source: &impl BlockValueSource,
+        source: &S,
         missing_opt: Option<u64>,
         ordered: bool,
     ) {
@@ -176,7 +155,7 @@ impl ColumnBlockAccessor {
     pub(crate) fn fetch_block_with_missing_unique_per_doc(
         &mut self,
         docs: &[DocId],
-        source: &impl BlockValueSource,
+        source: &dyn ValueSource,
         missing: Option<u64>,
         ordered: bool,
     ) {
@@ -288,37 +267,6 @@ impl ColumnBlockAccessor {
     }
 }
 
-#[inline]
-fn load_full_column_values(docs: &[DocId], accessor: &Column<u64>, values: &mut Vec<u64>) {
-    // Skip the resize when already the right length (common case: fixed-size blocks).
-    if values.len() != docs.len() {
-        values.resize(docs.len(), 0u64);
-    }
-    // When the docs form a contiguous ascending run we can fetch the values as a single range.
-    // This lets codecs (e.g. bitpacked) bulk-decode the slice instead of gathering value-by-value.
-    if is_contiguous(docs) {
-        accessor.values.get_range(docs[0] as u64, values);
-    } else {
-        accessor.values.get_vals(docs, values);
-    }
-}
-
-/// Returns true if `docs` is a contiguous ascending run `[d, d + 1, ..., d + n - 1]`.
-///
-/// Assumes `docs` is sorted ascending and free of duplicates (the invariant for the
-/// doc blocks passed to `fetch_block`), so comparing the endpoints is sufficient.
-#[inline]
-fn is_contiguous(docs: &[u32]) -> bool {
-    let (Some(&first), Some(&last)) = (docs.first(), docs.last()) else {
-        return false;
-    };
-    debug_assert!(
-        docs.windows(2).all(|w| w[0] < w[1]),
-        "fetch_block requires docs sorted ascending without duplicates"
-    );
-    (last - first) as usize + 1 == docs.len()
-}
-
 /// Given two sorted lists of docids `docs` and `hits`, hits is a subset of `docs`.
 /// Write in the output Vec all of the docs that are not in `hits`.
 ///
@@ -356,14 +304,23 @@ fn find_missing_docs(docs: &[u32], hits: &[u32], output: &mut Vec<u32>) {
 #[cfg(test)]
 #[allow(clippy::field_reassign_with_default)]
 mod tests {
+    use std::sync::Arc;
+
+    use columnar::{Column, ColumnType};
+
     use super::*;
 
+    #[derive(Debug)]
     struct TestValueSource {
         cardinality: Cardinality,
         entries: Vec<(DocId, u64)>,
     }
 
-    impl BlockValueSource for TestValueSource {
+    impl ValueSource for TestValueSource {
+        fn column_type(&self) -> ColumnType {
+            ColumnType::U64
+        }
+
         fn load_block(
             &self,
             docs: &[DocId],
@@ -385,6 +342,75 @@ mod tests {
         }
     }
 
+    #[test]
+    fn test_fetch_block_accepts_trait_object() {
+        let docs = [2, 4, 8];
+        let source = TestValueSource {
+            cardinality: Cardinality::Full,
+            entries: vec![(2, 20), (4, 40), (8, 80)],
+        };
+        let dyn_source: &dyn ValueSource = &source;
+        let mut accessor = ColumnBlockAccessor::default();
+
+        accessor.fetch_block(&docs, dyn_source);
+
+        assert!(accessor.has_one_value_per_doc(&docs));
+        assert_eq!(
+            accessor.iter_docid_vals(&docs).collect::<Vec<_>>(),
+            [(2, 20), (4, 40), (8, 80)]
+        );
+    }
+
+    fn full_column(vals: &[u64]) -> Column<u64> {
+        use columnar::column_index::ColumnIndex;
+        use columnar::column_values::{
+            serialize_and_load_u64_based_column_values, ALL_U64_CODEC_TYPES,
+        };
+        Column {
+            index: ColumnIndex::Full,
+            values: serialize_and_load_u64_based_column_values::<u64>(&vals, &ALL_U64_CODEC_TYPES),
+        }
+    }
+
+    #[test]
+    fn test_as_column_distinguishes_the_two_kinds() {
+        let column: Arc<dyn ValueSource> = Arc::new((full_column(&[5, 6, 7]), ColumnType::U64));
+        assert!(column.as_column().is_some());
+        assert_eq!(column.bounds(), Some((5, 7)));
+
+        let computed: Arc<dyn ValueSource> = Arc::new(TestValueSource {
+            cardinality: Cardinality::Full,
+            entries: vec![(0, 1)],
+        });
+        assert!(computed.as_column().is_none());
+        // No global view of a computed source, so no bounds and no bounds-driven fast paths.
+        assert_eq!(computed.bounds(), None);
+    }
+
+    #[test]
+    fn test_fetch_source_block_with_missing_on_a_computed_source() {
+        // A computed source reports what it produced: docs 0 and 2 have no value, so they are
+        // absent from `docids` and the source is `Optional`.
+        let docs = [0, 1, 2, 3];
+        let computed: Arc<dyn ValueSource> = Arc::new(TestValueSource {
+            cardinality: Cardinality::Optional,
+            entries: vec![(1, 11), (3, 33)],
+        });
+        let mut accessor = ColumnBlockAccessor::default();
+
+        accessor.fetch_block(&docs, &*computed);
+        assert!(!accessor.has_one_value_per_doc(&docs));
+        assert_eq!(
+            accessor.iter_docid_vals(&docs).collect::<Vec<_>>(),
+            [(1, 11), (3, 33)]
+        );
+
+        accessor.fetch_block_with_missing(&docs, &*computed, Some(99));
+        let mut pairs = accessor.iter_docid_vals(&docs).collect::<Vec<_>>();
+        pairs.sort_unstable();
+        assert_eq!(pairs, [(0, 99), (1, 11), (2, 99), (3, 33)]);
+    }
+
     #[test]
     fn test_find_missing_docs() {
         let docs: Vec<u32> = vec![1, 2, 3, 4, 5, 6, 7, 8, 9, 10];
@@ -392,30 +418,25 @@ mod tests {
 
         let mut missing_docs: Vec<u32> = Vec::new();
         find_missing_docs(&docs, &hits, &mut missing_docs);
-        assert_eq!(missing_docs, vec![1, 3, 5, 7, 9]);
+        assert_eq!(missing_docs, [1, 3, 5, 7, 9]);
     }
 
     #[test]
     fn test_find_missing_docs_empty() {
         let docs: Vec<u32> = Vec::new();
         let hits: Vec<u32> = vec![2, 4, 6, 8, 10];
-
         let mut missing_docs: Vec<u32> = Vec::new();
-
         find_missing_docs(&docs, &hits, &mut missing_docs);
-
-        assert_eq!(missing_docs, Vec::<u32>::new());
+        assert_eq!(missing_docs, [0u32; 0]);
     }
 
     #[test]
     fn test_find_missing_docs_all_missing() {
-        let docs: Vec<u32> = vec![1, 2, 3, 4, 5];
-        let hits: Vec<u32> = Vec::new();
-
+        let docs: &[u32] = &[1, 2, 3, 4, 5];
+        let hits: &[u32] = &[];
         let mut missing_docs: Vec<u32> = vec![10];
-        find_missing_docs(&docs, &hits, &mut missing_docs);
-
-        assert_eq!(missing_docs, vec![1, 2, 3, 4, 5]);
+        find_missing_docs(docs, hits, &mut missing_docs);
+        assert_eq!(&missing_docs, &[1u32, 2, 3, 4, 5]);
     }
 
     #[test]
@@ -426,13 +447,11 @@ mod tests {
             entries: vec![(2, 20), (4, 40), (8, 80)],
         };
         let mut accessor = ColumnBlockAccessor::default();
-
         accessor.fetch_block(&docs, &source);
-
         assert!(accessor.has_one_value_per_doc(&docs));
         assert_eq!(
             accessor.iter_docid_vals(&docs).collect::<Vec<_>>(),
-            vec![(2, 20), (4, 40), (8, 80)]
+            [(2, 20), (4, 40), (8, 80)]
         );
     }
 
@@ -450,7 +469,7 @@ mod tests {
         assert!(accessor.has_one_value_per_doc(&docs));
         assert_eq!(
             accessor.iter_docid_vals(&docs).collect::<Vec<_>>(),
-            vec![(0, 99), (1, 10), (2, 99), (4, 40)]
+            [(0, 99), (1, 10), (2, 99), (4, 40)]
         );
     }
 
@@ -468,7 +487,7 @@ mod tests {
         assert!(!accessor.has_one_value_per_doc(&docs));
         assert_eq!(
             accessor.iter_docid_vals(&docs).collect::<Vec<_>>(),
-            vec![(0, 1), (0, 3), (1, 5)]
+            [(0, 1), (0, 3), (1, 5)]
         );
     }
 
@@ -489,15 +508,20 @@ mod tests {
         let docs = [0, 1, 2, 4, 7, 8];
         let mut accessor = ColumnBlockAccessor::default();
 
-        accessor.fetch_block_with_missing_ordered(&docs, &column, Some(99), true);
+        accessor.fetch_block_with_missing_ordered(
+            &docs,
+            &(&column, ColumnType::U64),
+            Some(99),
+            true,
+        );
 
         assert_eq!(
             accessor.iter_vals().collect::<Vec<_>>(),
-            vec![99, 10, 99, 40, 70, 99]
+            [99, 10, 99, 40, 70, 99]
         );
         assert_eq!(
             accessor.iter_docid_vals(&docs).collect::<Vec<_>>(),
-            vec![(0, 99), (1, 10), (2, 99), (4, 40), (7, 70), (8, 99)]
+            [(0, 99), (1, 10), (2, 99), (4, 40), (7, 70), (8, 99)]
         );
     }
 
@@ -507,8 +531,8 @@ mod tests {
         accessor.docid_cache = vec![0, 0, 2, 3];
         accessor.val_cache = vec![10, 10, 10, 10];
         accessor.dedup_docid_val_pairs();
-        assert_eq!(accessor.docid_cache, vec![0, 2, 3]);
-        assert_eq!(accessor.val_cache, vec![10, 10, 10]);
+        assert_eq!(accessor.docid_cache, [0, 2, 3]);
+        assert_eq!(accessor.val_cache, [10, 10, 10]);
     }
 
     #[test]
@@ -518,8 +542,8 @@ mod tests {
         accessor.docid_cache = vec![0, 0, 0];
         accessor.val_cache = vec![1, 2, 1];
         accessor.dedup_docid_val_pairs();
-        assert_eq!(accessor.docid_cache, vec![0, 0]);
-        assert_eq!(accessor.val_cache, vec![1, 2]);
+        assert_eq!(accessor.docid_cache, [0, 0]);
+        assert_eq!(accessor.val_cache, [1, 2]);
     }
 
     #[test]
@@ -529,8 +553,8 @@ mod tests {
         accessor.docid_cache = vec![0, 0, 0, 1, 1];
         accessor.val_cache = vec![3, 1, 3, 5, 5];
         accessor.dedup_docid_val_pairs();
-        assert_eq!(accessor.docid_cache, vec![0, 0, 1]);
-        assert_eq!(accessor.val_cache, vec![1, 3, 5]);
+        assert_eq!(accessor.docid_cache, [0, 0, 1]);
+        assert_eq!(accessor.val_cache, [1, 3, 5]);
     }
 
     #[test]
@@ -539,8 +563,8 @@ mod tests {
         accessor.docid_cache = vec![0, 0, 1];
         accessor.val_cache = vec![1, 2, 3];
         accessor.dedup_docid_val_pairs();
-        assert_eq!(accessor.docid_cache, vec![0, 0, 1]);
-        assert_eq!(accessor.val_cache, vec![1, 2, 3]);
+        assert_eq!(accessor.docid_cache, [0, 0, 1]);
+        assert_eq!(accessor.val_cache, [1, 2, 3]);
     }
 
     #[test]
@@ -549,18 +573,8 @@ mod tests {
         accessor.docid_cache = vec![0];
         accessor.val_cache = vec![1];
         accessor.dedup_docid_val_pairs();
-        assert_eq!(accessor.docid_cache, vec![0]);
-        assert_eq!(accessor.val_cache, vec![1]);
-    }
-
-    #[test]
-    fn test_is_contiguous() {
-        assert!(!is_contiguous(&[]));
-        assert!(is_contiguous(&[5]));
-        assert!(is_contiguous(&[5, 6, 7, 8]));
-        assert!(is_contiguous(&[0, 1, 2]));
-        assert!(!is_contiguous(&[5, 7, 8]));
-        assert!(!is_contiguous(&[0, 1, 3]));
+        assert_eq!(accessor.docid_cache, [0]);
+        assert_eq!(accessor.val_cache, [1]);
     }
 
     #[test]
@@ -579,7 +593,7 @@ mod tests {
         };
 
         let check = |accessor: &mut ColumnBlockAccessor, docs: &[u32]| {
-            accessor.fetch_block(docs, &column);
+            accessor.fetch_block(docs, &(&column, ColumnType::U64));
             let got: Vec<(u32, u64)> = accessor.iter_docid_vals(docs).collect();
             let expected: Vec<(u32, u64)> = docs.iter().map(|&d| (d, vals[d as usize])).collect();
             assert_eq!(got, expected);
diff --git a/src/aggregation/value_source/mod.rs b/src/aggregation/value_source/mod.rs
new file mode 100644
index 000000000..12fec6a41
--- /dev/null
+++ b/src/aggregation/value_source/mod.rs
@@ -0,0 +1,126 @@
+mod block_accessor;
+mod value_source_registry;
+
+#[cfg(test)]
+pub(crate) mod tests;
+
+use std::borrow::Borrow;
+
+pub(crate) use block_accessor::ColumnBlockAccessor;
+use columnar::{Cardinality, Column, ColumnType, ColumnValues, RowId};
+pub use value_source_registry::{ValueSourceProvider, ValueSourceRegistry};
+
+use crate::DocId;
+
+/// A source of values for a block of documents.
+pub trait ValueSource: std::fmt::Debug {
+    /// Logical type of the encoded values returned by this source.
+    ///
+    /// Numeric values use the corresponding monotonic `u64` mapping; string/bytes
+    /// values are dictionary ordinals and IP addresses use the compact space ord.
+    fn column_type(&self) -> ColumnType;
+
+    /// Loads the values for `docs` into `values`.
+    ///
+    /// Precondition: `docs` has to be strictly increasing.
+    ///
+    /// The output buffers are reused across blocks: on entry, `values` and `docids`
+    /// hold stale data from a previous call. Implementations must clear their
+    /// content (not append to it).
+    ///
+    /// On return, depending on the returned `Cardinality`:
+    /// - `Full`: `values.len() == docs.len()` and `values[i]` is the value of `docs[i]`. `docids`
+    ///   is left unspecified and must not be read by the caller.
+    /// - `Optional` / `Multivalued`: `docids.len() == values.len()` and `values[i]` is a value of
+    ///   `docids[i]`. `docids` only contains docs from `docs`. A doc is repeated once per value.
+    ///
+    /// `row_ids` is scratch the implementation may use freely.
+    fn load_block(
+        &self,
+        docs: &[DocId],
+        values: &mut Vec<u64>,
+        docids: &mut Vec<DocId>,
+        row_ids: &mut Vec<RowId>,
+    ) -> Cardinality;
+
+    /// Returns the physical column, if this source is backed by one.
+    fn as_column(&self) -> Option<&Column<u64>> {
+        None
+    }
+
+    /// Global value bounds, for fast paths that need to size or clamp something up front.
+    fn bounds(&self) -> Option<(u64, u64)> {
+        let column = self.as_column()?;
+        Some((column.min_value(), column.max_value()))
+    }
+}
+
+// Lenient columns have erased their logical type; the tuple retains it alongside the values.
+impl<ColumnRef: Borrow<Column<u64>> + std::fmt::Debug> ValueSource for (ColumnRef, ColumnType) {
+    #[inline]
+    fn column_type(&self) -> ColumnType {
+        self.1
+    }
+
+    #[inline]
+    fn load_block(
+        &self,
+        docs: &[DocId],
+        values: &mut Vec<u64>,
+        docids: &mut Vec<DocId>,
+        row_ids: &mut Vec<RowId>,
+    ) -> Cardinality {
+        let column = self.0.borrow();
+        let cardinality = column.index.get_cardinality();
+        if cardinality.is_full() {
+            load_full_column_values(docs, &*column.values, values);
+        } else {
+            docids.clear();
+            row_ids.clear();
+            column.row_ids_for_docs(docs, docids, row_ids);
+            values.resize(row_ids.len(), 0u64);
+            column.values.get_vals(row_ids, values);
+        }
+        cardinality
+    }
+
+    #[inline]
+    fn as_column(&self) -> Option<&Column<u64>> {
+        Some(self.0.borrow())
+    }
+}
+
+/// `docs` has to be sorted ascending and free of duplicates.
+#[inline]
+fn load_full_column_values(
+    docs: &[DocId],
+    column_values: &dyn ColumnValues<u64>,
+    values: &mut Vec<u64>,
+) {
+    // Skip the resize when already the right length (common case: fixed-size blocks).
+    if values.len() != docs.len() {
+        values.resize(docs.len(), 0u64);
+    }
+    // When the docs form a contiguous ascending run we can fetch the values as a single range.
+    // This lets codecs (e.g. bitpacked) bulk-decode the slice instead of gathering value-by-value.
+    if is_contiguous(docs) {
+        column_values.get_range(docs[0] as u64, values);
+    } else {
+        column_values.get_vals(docs, values);
+    }
+}
+
+/// Returns true if `docs` is a contiguous ascending run `[d, d + 1, ..., d + n - 1]`.
+///
+/// `docs` has to be sorted ascending and free of duplicates.
+#[inline]
+fn is_contiguous(docs: &[u32]) -> bool {
+    let (Some(&first), Some(&last)) = (docs.first(), docs.last()) else {
+        return false;
+    };
+    debug_assert!(
+        docs.windows(2).all(|w| w[0] < w[1]),
+        "fetch_block requires docs sorted ascending without duplicates"
+    );
+    (last - first) as usize + 1 == docs.len()
+}
diff --git a/src/aggregation/value_source/tests.rs b/src/aggregation/value_source/tests.rs
new file mode 100644
index 000000000..b2a9c3b4f
--- /dev/null
+++ b/src/aggregation/value_source/tests.rs
@@ -0,0 +1,117 @@
+use std::sync::Arc;
+
+use columnar::ColumnType;
+
+use super::*;
+use crate::SegmentReader;
+
+#[derive(Debug)]
+pub(crate) struct Constant(u64);
+
+impl ValueSource for Constant {
+    fn column_type(&self) -> ColumnType {
+        ColumnType::U64
+    }
+
+    fn load_block(
+        &self,
+        docs: &[DocId],
+        values: &mut Vec<u64>,
+        _docids: &mut Vec<DocId>,
+        _row_ids: &mut Vec<columnar::RowId>,
+    ) -> Cardinality {
+        values.clear();
+        values.resize(docs.len(), self.0);
+        Cardinality::Full
+    }
+}
+
+pub(crate) struct ConstantProvider(pub u64);
+
+impl ValueSourceProvider for ConstantProvider {
+    fn for_segment(&self, _reader: &SegmentReader) -> crate::Result<Arc<dyn ValueSource>> {
+        Ok(Arc::new(Constant(self.0)))
+    }
+}
+fn index_with_scores(scores: &[u64]) -> crate::Index {
+    use crate::schema::{Schema, FAST};
+    let mut builder = Schema::builder();
+    let score = builder.add_u64_field("score", FAST);
+    let index = crate::Index::create_in_ram(builder.build());
+    let mut writer = index.writer_for_tests().unwrap();
+    for &value in scores {
+        writer.add_document(crate::doc!(score => value)).unwrap();
+    }
+    writer.commit().unwrap();
+    index
+}
+
+fn run_agg(index: &crate::Index, aggs: serde_json::Value) -> serde_json::Value {
+    let mut registry = ValueSourceRegistry::default();
+    registry.register("computed", Arc::new(ConstantProvider(1u64)));
+    run_agg_with_registry(index, aggs, registry)
+}
+
+fn run_agg_with_registry(
+    index: &crate::Index,
+    aggs: serde_json::Value,
+    registry: ValueSourceRegistry,
+) -> serde_json::Value {
+    use crate::aggregation::agg_req::Aggregations;
+    use crate::aggregation::{AggContextParams, AggregationCollector};
+    use crate::query::AllQuery;
+
+    let context = AggContextParams::default().with_value_sources(Arc::new(registry));
+    let aggs: Aggregations = serde_json::from_value(aggs).unwrap();
+    let collector = AggregationCollector::from_aggs(aggs, context);
+    let searcher = index.reader().unwrap().searcher();
+    let result = searcher.search(&AllQuery, &collector).unwrap();
+    serde_json::to_value(result).unwrap()
+}
+
+#[test]
+fn test_metric_over_registered_source() {
+    let index = index_with_scores(&[10, 20, 30, 40]);
+    let result = run_agg(
+        &index,
+        serde_json::json!({ "s": { "stats": { "field": "computed" } } }),
+    );
+    // Every document contributes exactly 1.
+    assert_eq!(result["s"]["count"], 4);
+    assert_eq!(result["s"]["sum"], 4.0);
+    assert_eq!(result["s"]["avg"], 1.0);
+    assert_eq!(result["s"]["min"], 1.0);
+    assert_eq!(result["s"]["max"], 1.0);
+}
+
+#[test]
+fn test_registered_source_as_sub_aggregation_of_terms() {
+    let index = index_with_scores(&[7, 7, 7, 9]);
+    let result = run_agg(
+        &index,
+        serde_json::json!({
+            "by_score": {
+                "terms": { "field": "score" },
+                "aggs": { "s": { "sum": { "field": "computed" } } }
+            }
+        }),
+    );
+    let buckets = result["by_score"]["buckets"].as_array().unwrap();
+    assert_eq!(buckets.len(), 2);
+    assert_eq!(buckets[0]["key"], 7.0);
+    assert_eq!(buckets[0]["doc_count"], 3);
+    assert_eq!(buckets[0]["s"]["value"], 3.0);
+    assert_eq!(buckets[1]["key"], 9.0);
+    assert_eq!(buckets[1]["doc_count"], 1);
+    assert_eq!(buckets[1]["s"]["value"], 1.0);
+}
+
+#[test]
+fn test_is_contiguous() {
+    assert!(!is_contiguous(&[]));
+    assert!(is_contiguous(&[5]));
+    assert!(is_contiguous(&[5, 6, 7, 8]));
+    assert!(is_contiguous(&[0, 1, 2]));
+    assert!(!is_contiguous(&[5, 7, 8]));
+    assert!(!is_contiguous(&[0, 1, 3]));
+}
diff --git a/src/aggregation/value_source/value_source_registry.rs b/src/aggregation/value_source/value_source_registry.rs
new file mode 100644
index 000000000..f9888a4e6
--- /dev/null
+++ b/src/aggregation/value_source/value_source_registry.rs
@@ -0,0 +1,72 @@
+//! Registration of named, computed value sources.
+
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use super::ValueSource;
+use crate::SegmentReader;
+
+/// Creates a value source for each segment.
+pub trait ValueSourceProvider: Send + Sync + 'static {
+    /// Binds this definition to a single segment.
+    fn for_segment(&self, reader: &SegmentReader) -> crate::Result<Arc<dyn ValueSource>>;
+}
+
+/// Named computed sources available to an aggregation request.
+#[derive(Clone, Default)]
+pub struct ValueSourceRegistry {
+    providers: HashMap<String, Arc<dyn ValueSourceProvider>>,
+}
+
+impl ValueSourceRegistry {
+    /// Registers `provider` under `name`, which aggregation requests then use as a field name.
+    ///
+    /// Inserting the same name several times results in an override.
+    pub fn register(&mut self, name: &str, provider: Arc<dyn ValueSourceProvider>) {
+        let name = name.to_string();
+        self.providers.insert(name, provider);
+    }
+
+    #[inline]
+    pub(crate) fn get(&self, name: &str) -> Option<&Arc<dyn ValueSourceProvider>> {
+        self.providers.get(name)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::aggregation::value_source::tests::ConstantProvider;
+    use crate::schema::Schema;
+
+    #[test]
+    fn test_register_then_get() {
+        let mut registry = ValueSourceRegistry::default();
+        registry.register("computed", Arc::new(ConstantProvider(1)));
+        assert!(registry.get("computed").is_some());
+        assert!(registry.get("absent").is_none());
+    }
+
+    #[test]
+    fn test_register_overrides() {
+        let mut registry = ValueSourceRegistry::default();
+        registry.register("computed", Arc::new(ConstantProvider(1)));
+        registry.register("computed", Arc::new(ConstantProvider(2)));
+        let index = crate::Index::create_in_ram(Schema::builder().build());
+        let mut writer = index.writer_for_tests().unwrap();
+        writer.add_document(crate::doc!()).unwrap();
+        writer.commit().unwrap();
+        let searcher = index.reader().unwrap().searcher();
+        let value_source_provider = registry.get("computed").unwrap();
+        let value_source = value_source_provider
+            .for_segment(searcher.segment_reader(0u32))
+            .unwrap();
+        let mut values = Vec::new();
+        let mut doc_ids = Vec::new();
+        let mut row_ids = Vec::new();
+        let docs = &[1u32];
+        value_source.load_block(docs, &mut values, &mut doc_ids, &mut row_ids);
+        assert!(doc_ids.is_empty());
+        assert_eq!(&values, &[2u64]);
+    }
+}
diff --git a/src/directory/composite_file.rs b/src/directory/composite_file.rs
index 93e063880..f3d3d4cd6 100644
--- a/src/directory/composite_file.rs
+++ b/src/directory/composite_file.rs
@@ -183,7 +183,6 @@ impl CompositeFile {
 #[cfg(test)]
 mod test {
 
-    use std::io::Write;
     use std::path::Path;
 
     use common::{BinarySerializable, VInt};
@@ -201,10 +200,8 @@ mod test {
             let mut composite_write = CompositeWrite::wrap(w);
             let mut write_0 = composite_write.for_field(Field::from_field_id(0u32));
             VInt(32431123u64).serialize(&mut write_0)?;
-            write_0.flush()?;
             let mut write_4 = composite_write.for_field(Field::from_field_id(4u32));
             VInt(2).serialize(&mut write_4)?;
-            write_4.flush()?;
             composite_write.close()?;
         }
         {
@@ -243,13 +240,10 @@ mod test {
             let mut composite_write = CompositeWrite::wrap(w);
             let mut write = composite_write.for_field_with_idx(Field::from_field_id(1u32), 0);
             VInt(32431123u64).serialize(&mut write)?;
-            write.flush()?;
-            let write = composite_write.for_field_with_idx(Field::from_field_id(1u32), 1);
-            write.flush()?;
+            composite_write.for_field_with_idx(Field::from_field_id(1u32), 1);
 
             let mut write = composite_write.for_field_with_idx(Field::from_field_id(0u32), 0);
             VInt(1_000_000).serialize(&mut write)?;
-            write.flush()?;
 
             composite_write.close()?;
         }
diff --git a/src/directory/directory.rs b/src/directory/directory.rs
index 0bc4b7f95..9e3a7b14d 100644
--- a/src/directory/directory.rs
+++ b/src/directory/directory.rs
@@ -1,4 +1,3 @@
-use std::io::Write;
 use std::path::{Path, PathBuf};
 use std::sync::Arc;
 use std::time::Duration;
@@ -6,7 +5,9 @@ use std::{fmt, io, thread};
 
 use crate::directory::directory_lock::Lock;
 use crate::directory::error::{DeleteError, LockError, OpenReadError, OpenWriteError};
-use crate::directory::{FileHandle, FileSlice, WatchCallback, WatchHandle, WritePtr};
+use crate::directory::{
+    FileHandle, FileSlice, TerminatingWrite, WatchCallback, WatchHandle, WritePtr,
+};
 
 /// Retry the logic of acquiring locks is pretty simple.
 /// We just retry `n` times after a given `duratio`, both
@@ -75,11 +76,11 @@ fn try_acquire_lock(
     filepath: &Path,
     directory: &dyn Directory,
 ) -> Result<DirectoryLock, TryAcquireLockError> {
-    let mut write = directory.open_write(filepath).map_err(|e| match e {
+    let write = directory.open_write(filepath).map_err(|e| match e {
         OpenWriteError::FileAlreadyExists(_) => TryAcquireLockError::FileExists,
         OpenWriteError::IoError { io_error, .. } => TryAcquireLockError::IoError(io_error),
     })?;
-    write.flush().map_err(TryAcquireLockError::from)?;
+    write.terminate().map_err(TryAcquireLockError::from)?;
     Ok(DirectoryLock::from(Box::new(DirectoryLockGuard {
         directory: directory.box_clone(),
         path: filepath.to_owned(),
@@ -138,10 +139,8 @@ pub trait Directory: DirectoryClone + fmt::Debug + Send + Sync + 'static {
     /// Opens a writer for the *virtual file* associated with
     /// a [`Path`].
     ///
-    /// Right after this call, for the span of the execution of the program
-    /// the file should be created and any subsequent call to
-    /// [`Directory::open_read()`] for the same path should return
-    /// a [`FileSlice`].
+    /// After the writer is terminated, the file should be created and any subsequent call to
+    /// [`Directory::open_read()`] for the same path should return a [`FileSlice`].
     ///
     /// However, depending on the directory implementation,
     /// it might be required to call [`Directory::sync_directory()`] to ensure
@@ -150,15 +149,11 @@ pub trait Directory: DirectoryClone + fmt::Debug + Send + Sync + 'static {
     /// a POSIX filesystem.)
     ///
     /// Write operations may be aggressively buffered.
-    /// The client of this trait is responsible for calling flush
+    /// The client of this trait is responsible for calling terminate
     /// to ensure that subsequent `read` operations
     /// will take into account preceding `write` operations.
     ///
-    /// Flush operation should also be persistent.
-    ///
-    /// The user shall not rely on [`Drop`] triggering `flush`.
-    /// Note that [`RamDirectory`][crate::directory::RamDirectory] will
-    /// panic! if `flush` was not called.
+    /// The user shall not rely on [`Drop`] triggering terminate.
     ///
     /// The file may not previously exist.
     fn open_write(&self, path: &Path) -> Result<WritePtr, OpenWriteError>;
diff --git a/src/directory/mmap_directory/mod.rs b/src/directory/mmap_directory/mod.rs
index 0370a8b9e..1c33f0358 100644
--- a/src/directory/mmap_directory/mod.rs
+++ b/src/directory/mmap_directory/mod.rs
@@ -319,7 +319,7 @@ impl Drop for ReleaseLockFile {
 }
 
 /// This Write wraps a File, but has the specificity of
-/// call `sync_all` on flush.
+/// calling `sync_all` on terminate.
 struct SafeFileWriter(File);
 
 impl SafeFileWriter {
@@ -432,7 +432,7 @@ impl Directory for MmapDirectory {
             .create_new(true)
             .open(full_path);
 
-        let mut file = open_res.map_err(|io_err| {
+        let file = open_res.map_err(|io_err| {
             if io_err.kind() == io::ErrorKind::AlreadyExists {
                 OpenWriteError::FileAlreadyExists(path.to_path_buf())
             } else {
@@ -440,10 +440,6 @@ impl Directory for MmapDirectory {
             }
         })?;
 
-        // making sure the file is created.
-        file.flush()
-            .map_err(|io_error| OpenWriteError::wrap_io_error(io_error, path.to_path_buf()))?;
-
         // Note we actually do not sync the parent directory here.
         //
         // A newly created file, may, in some case, be created and even flushed to disk.
@@ -559,10 +555,11 @@ mod tests {
         // In that case the directory returns a SharedVecSlice.
         let mmap_directory = MmapDirectory::create_from_tempdir().unwrap();
         let path = PathBuf::from("test");
-        {
-            let mut w = mmap_directory.open_write(&path).unwrap();
-            w.flush().unwrap();
-        }
+        mmap_directory
+            .open_write(&path)
+            .unwrap()
+            .terminate()
+            .unwrap();
         let readonlymap = mmap_directory.open_read(&path).unwrap();
         assert_eq!(readonlymap.len(), 0);
     }
@@ -578,12 +575,10 @@ mod tests {
         let paths: Vec<PathBuf> = (0..num_paths)
             .map(|i| PathBuf::from(&*format!("file_{i}")))
             .collect();
-        {
-            for path in &paths {
-                let mut w = mmap_directory.open_write(path).unwrap();
-                w.write_all(content).unwrap();
-                w.flush().unwrap();
-            }
+        for path in &paths {
+            let mut w = mmap_directory.open_write(path).unwrap();
+            w.write_all(content).unwrap();
+            w.terminate().unwrap();
         }
 
         let mut keep = vec![];
diff --git a/src/directory/ram_directory.rs b/src/directory/ram_directory.rs
index 7ea6db9e0..2ce9d39d1 100644
--- a/src/directory/ram_directory.rs
+++ b/src/directory/ram_directory.rs
@@ -1,6 +1,7 @@
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::io::{self, BufWriter, Cursor, Write};
 use std::path::{Path, PathBuf};
+use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, RwLock};
 use std::{fmt, result};
 
@@ -14,6 +15,49 @@ use crate::directory::{
     WatchHandle, WritePtr,
 };
 
+const MEMORY_USAGE_UPDATE_THRESHOLD: usize = 10_000;
+
+struct MemoryUsageTracker {
+    shared_usage: Arc<AtomicUsize>,
+    reported_bytes: usize,
+    unreported_bytes: usize,
+}
+
+impl MemoryUsageTracker {
+    fn new(shared_usage: Arc<AtomicUsize>) -> Self {
+        Self {
+            shared_usage,
+            reported_bytes: 0,
+            unreported_bytes: 0,
+        }
+    }
+
+    fn add(&mut self, num_bytes: usize) {
+        self.unreported_bytes += num_bytes;
+        if self.unreported_bytes >= MEMORY_USAGE_UPDATE_THRESHOLD {
+            self.shared_usage
+                .fetch_add(self.unreported_bytes, Ordering::Relaxed);
+            self.reported_bytes += self.unreported_bytes;
+            self.unreported_bytes = 0;
+        }
+    }
+
+    fn release(&mut self) {
+        if self.reported_bytes > 0 {
+            self.shared_usage
+                .fetch_sub(self.reported_bytes, Ordering::Relaxed);
+            self.reported_bytes = 0;
+        }
+        self.unreported_bytes = 0;
+    }
+}
+
+impl Drop for MemoryUsageTracker {
+    fn drop(&mut self) {
+        self.release();
+    }
+}
+
 /// Writer associated with the [`RamDirectory`].
 ///
 /// The Writer just writes a buffer.
@@ -21,64 +65,82 @@ struct VecWriter {
     path: PathBuf,
     shared_directory: RamDirectory,
     data: Cursor<Vec<u8>>,
-    is_flushed: bool,
+    memory_usage: MemoryUsageTracker,
+    is_finished: bool,
 }
 
 impl VecWriter {
     fn new(path_buf: PathBuf, shared_directory: RamDirectory) -> VecWriter {
+        let memory_usage =
+            MemoryUsageTracker::new(Arc::clone(&shared_directory.active_writer_mem_usage));
         VecWriter {
             path: path_buf,
             data: Cursor::new(Vec::new()),
             shared_directory,
-            is_flushed: true,
+            memory_usage,
+            is_finished: true,
         }
     }
 }
 
 impl Drop for VecWriter {
     fn drop(&mut self) {
-        if !self.is_flushed {
+        if !self.is_finished {
             warn!(
-                "You forgot to flush {:?} before its writer got Drop. Do not rely on drop. This \
-                 also occurs when the indexer crashed, so you may want to check the logs for the \
-                 root cause.",
+                "You forgot to terminate {:?} before its writer got Drop. Do not rely on drop. \
+                 This also occurs when the indexer crashed, so you may want to check the logs for \
+                 the root cause.",
                 self.path
-            )
+            );
+        }
+        if let Ok(mut fs) = self.shared_directory.fs.write() {
+            fs.active_writers.remove(&self.path);
         }
     }
 }
 
 impl Write for VecWriter {
     fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
-        self.is_flushed = false;
+        self.is_finished = false;
+        let previous_capacity = self.data.get_ref().capacity();
         self.data.write_all(buf)?;
+        let capacity = self.data.get_ref().capacity();
+        self.memory_usage.add(capacity - previous_capacity);
         Ok(buf.len())
     }
 
     fn flush(&mut self) -> io::Result<()> {
-        self.is_flushed = true;
-        let mut fs = self.shared_directory.fs.write().unwrap();
-        fs.write(self.path.clone(), self.data.get_ref());
         Ok(())
     }
 }
 
 impl TerminatingWrite for VecWriter {
     fn terminate_ref(&mut self, _: AntiCallToken) -> io::Result<()> {
-        self.flush()
+        let mut data = std::mem::take(self.data.get_mut());
+        data.shrink_to_fit();
+        let mut fs = self.shared_directory.fs.write().unwrap();
+        fs.active_writers.remove(&self.path);
+        fs.write_owned(self.path.clone(), data);
+        self.is_finished = true;
+        self.memory_usage.release();
+        Ok(())
     }
 }
 
 #[derive(Default)]
 struct InnerDirectory {
     fs: HashMap<PathBuf, FileSlice>,
+    active_writers: HashSet<PathBuf>,
     watch_router: WatchCallbackList,
 }
 
 impl InnerDirectory {
     fn write(&mut self, path: PathBuf, data: &[u8]) -> bool {
-        let data = FileSlice::from(data.to_vec());
-        self.fs.insert(path, data).is_some()
+        self.write_owned(path, data.to_vec())
+    }
+
+    fn write_owned(&mut self, path: PathBuf, data: Vec<u8>) -> bool {
+        self.fs.insert(path, FileSlice::from(data)).is_some()
     }
 
     fn open_read(&self, path: &Path) -> Result<FileSlice, OpenReadError> {
@@ -96,7 +158,7 @@ impl InnerDirectory {
     }
 
     fn exists(&self, path: &Path) -> bool {
-        self.fs.contains_key(path)
+        self.fs.contains_key(path) || self.active_writers.contains(path)
     }
 
     fn watch(&mut self, watch_handle: WatchCallback) -> WatchHandle {
@@ -117,10 +179,11 @@ impl fmt::Debug for RamDirectory {
 /// A Directory storing everything in anonymous memory.
 ///
 /// It is mainly meant for unit testing.
-/// Writes are only made visible upon flushing.
+/// Writes are only made visible upon terminating the writer.
 #[derive(Clone, Default)]
 pub struct RamDirectory {
     fs: Arc<RwLock<InnerDirectory>>,
+    active_writer_mem_usage: Arc<AtomicUsize>,
 }
 
 impl RamDirectory {
@@ -129,24 +192,12 @@ impl RamDirectory {
         Self::default()
     }
 
-    /// Deep clones the directory.
+    /// Returns the size of the files and an estimate of active writer allocations.
     ///
-    /// Ulterior writes on one of the copy
-    /// will not affect the other copy.
-    pub fn deep_clone(&self) -> RamDirectory {
-        let inner_clone = InnerDirectory {
-            fs: self.fs.read().unwrap().fs.clone(),
-            watch_router: Default::default(),
-        };
-        RamDirectory {
-            fs: Arc::new(RwLock::new(inner_clone)),
-        }
-    }
-
-    /// Returns the sum of the size of the different files
-    /// in the [`RamDirectory`].
+    /// Active writer allocations are reported in 10 kB increments.
     pub fn total_mem_usage(&self) -> usize {
         self.fs.read().unwrap().total_mem_usage()
+            + self.active_writer_mem_usage.load(Ordering::Relaxed)
     }
 
     /// Write a copy of all of the files saved in the [`RamDirectory`] in the target [`Directory`].
@@ -198,16 +249,14 @@ impl Directory for RamDirectory {
     }
 
     fn open_write(&self, path: &Path) -> Result<WritePtr, OpenWriteError> {
-        let mut fs = self.fs.write().unwrap();
         let path_buf = PathBuf::from(path);
-        let vec_writer = VecWriter::new(path_buf.clone(), self.clone());
-        let exists = fs.write(path_buf.clone(), &[]);
-        // force the creation of the file to mimic the MMap directory.
-        if exists {
-            Err(OpenWriteError::FileAlreadyExists(path_buf))
-        } else {
-            Ok(BufWriter::new(Box::new(vec_writer)))
+        let mut fs = self.fs.write().unwrap();
+        if fs.exists(path) || !fs.active_writers.insert(path_buf.clone()) {
+            return Err(OpenWriteError::FileAlreadyExists(path_buf));
         }
+        drop(fs);
+        let vec_writer = VecWriter::new(path_buf, self.clone());
+        Ok(BufWriter::new(Box::new(vec_writer)))
     }
 
     fn atomic_read(&self, path: &Path) -> Result<Vec<u8>, OpenReadError> {
@@ -244,7 +293,8 @@ mod tests {
     use std::io::Write;
     use std::path::Path;
 
-    use super::RamDirectory;
+    use super::{RamDirectory, MEMORY_USAGE_UPDATE_THRESHOLD};
+    use crate::directory::TerminatingWrite;
     use crate::Directory;
 
     #[test]
@@ -257,7 +307,7 @@ mod tests {
         assert!(directory.atomic_write(path_atomic, msg_atomic).is_ok());
         let mut wrt = directory.open_write(path_seq).unwrap();
         assert!(wrt.write_all(msg_seq).is_ok());
-        assert!(wrt.flush().is_ok());
+        assert!(wrt.terminate().is_ok());
         let directory_copy = RamDirectory::create();
         assert!(directory.persist(&directory_copy).is_ok());
         assert_eq!(directory_copy.atomic_read(path_atomic).unwrap(), msg_atomic);
@@ -265,21 +315,45 @@ mod tests {
     }
 
     #[test]
-    fn test_ram_directory_deep_clone() {
+    fn test_dropped_writer_releases_path() {
         let dir = RamDirectory::default();
-        let test = Path::new("test");
-        let test2 = Path::new("test2");
-        dir.atomic_write(test, b"firstwrite").unwrap();
-        let dir_clone = dir.deep_clone();
-        assert_eq!(
-            dir_clone.atomic_read(test).unwrap(),
-            dir.atomic_read(test).unwrap()
-        );
-        dir.atomic_write(test, b"original").unwrap();
-        dir_clone.atomic_write(test, b"clone").unwrap();
-        dir_clone.atomic_write(test2, b"clone2").unwrap();
-        assert_eq!(dir.atomic_read(test).unwrap(), b"original");
-        assert_eq!(&dir_clone.atomic_read(test).unwrap(), b"clone");
-        assert_eq!(&dir_clone.atomic_read(test2).unwrap(), b"clone2");
+        let path = Path::new("file");
+        let writer = dir.open_write(path).unwrap();
+        assert!(dir.open_write(path).is_err());
+
+        drop(writer);
+        dir.open_write(path).unwrap().terminate().unwrap();
+        assert!(dir.exists(path).unwrap());
+    }
+
+    #[test]
+    fn test_active_writer_memory_usage() {
+        let dir = RamDirectory::default();
+        let path = Path::new("file");
+        let mut writer = dir.open_write(path).unwrap();
+        assert_eq!(dir.total_mem_usage(), 0);
+
+        writer.write_all(&[0u8]).unwrap();
+        assert_eq!(dir.total_mem_usage(), 0);
+        writer
+            .write_all(&vec![0u8; MEMORY_USAGE_UPDATE_THRESHOLD])
+            .unwrap();
+        let first_capacity = dir.total_mem_usage();
+        assert!(first_capacity >= MEMORY_USAGE_UPDATE_THRESHOLD);
+
+        writer.write_all(&vec![0u8; first_capacity + 1]).unwrap();
+        let grown_capacity = dir.total_mem_usage();
+        assert!(grown_capacity > first_capacity);
+
+        writer.flush().unwrap();
+        assert_eq!(dir.total_mem_usage(), grown_capacity);
+        assert_eq!(dir.clone().total_mem_usage(), dir.total_mem_usage());
+
+        let file_len = 1 + MEMORY_USAGE_UPDATE_THRESHOLD + first_capacity + 1;
+        writer.terminate().unwrap();
+        assert_eq!(dir.total_mem_usage(), file_len);
+
+        dir.delete(path).unwrap();
+        assert_eq!(dir.total_mem_usage(), 0);
     }
 }
diff --git a/src/directory/tests.rs b/src/directory/tests.rs
index a2c8473ce..dd6d5de6a 100644
--- a/src/directory/tests.rs
+++ b/src/directory/tests.rs
@@ -120,11 +120,10 @@ mod ram_directory_tests {
 fn test_simple(directory: &dyn Directory) -> crate::Result<()> {
     let test_path: &'static Path = Path::new("some_path_for_test");
     let mut write_file = directory.open_write(test_path)?;
-    assert!(directory.exists(test_path).unwrap());
     write_file.write_all(&[4])?;
     write_file.write_all(&[3])?;
     write_file.write_all(&[7, 3, 5])?;
-    write_file.flush()?;
+    write_file.terminate()?;
     let read_file = directory.open_read(test_path)?.read_bytes()?;
     assert_eq!(read_file.as_slice(), &[4u8, 3u8, 7u8, 3u8, 5u8]);
     mem::drop(read_file);
@@ -135,7 +134,9 @@ fn test_simple(directory: &dyn Directory) -> crate::Result<()> {
 
 fn test_rewrite_forbidden(directory: &dyn Directory) -> crate::Result<()> {
     let test_path: &'static Path = Path::new("some_path_for_test");
-    directory.open_write(test_path)?;
+    let writer = directory.open_write(test_path)?;
+    assert!(directory.open_write(test_path).is_err());
+    writer.terminate()?;
     assert!(directory.exists(test_path).unwrap());
     assert!(directory.open_write(test_path).is_err());
     assert!(directory.delete(test_path).is_ok());
@@ -144,13 +145,12 @@ fn test_rewrite_forbidden(directory: &dyn Directory) -> crate::Result<()> {
 
 fn test_write_create_the_file(directory: &dyn Directory) {
     let test_path: &'static Path = Path::new("some_path_for_test");
-    {
-        assert!(directory.open_read(test_path).is_err());
-        let _w = directory.open_write(test_path).unwrap();
-        assert!(directory.exists(test_path).unwrap());
-        assert!(directory.open_read(test_path).is_ok());
-        assert!(directory.delete(test_path).is_ok());
-    }
+    assert!(directory.open_read(test_path).is_err());
+    let writer = directory.open_write(test_path).unwrap();
+    assert!(directory.exists(test_path).unwrap());
+    writer.terminate().unwrap();
+    assert!(directory.open_read(test_path).is_ok());
+    assert!(directory.delete(test_path).is_ok());
 }
 
 fn test_directory_delete(directory: &dyn Directory) -> crate::Result<()> {
@@ -158,7 +158,7 @@ fn test_directory_delete(directory: &dyn Directory) -> crate::Result<()> {
     assert!(directory.open_read(test_path).is_err());
     let mut write_file = directory.open_write(test_path)?;
     write_file.write_all(&[1, 2, 3, 4])?;
-    write_file.flush()?;
+    write_file.terminate()?;
     {
         let read_handle = directory.open_read(test_path)?.read_bytes()?;
         assert_eq!(read_handle.as_slice(), &[1u8, 2u8, 3u8, 4u8]);
diff --git a/src/fieldnorm/serializer.rs b/src/fieldnorm/serializer.rs
index 316b4cfad..5326ea494 100644
--- a/src/fieldnorm/serializer.rs
+++ b/src/fieldnorm/serializer.rs
@@ -22,7 +22,6 @@ impl FieldNormsSerializer {
     pub fn serialize_field(&mut self, field: Field, fieldnorms_data: &[u8]) -> io::Result<()> {
         let write = self.composite_write.for_field(field);
         write.write_all(fieldnorms_data)?;
-        write.flush()?;
         Ok(())
     }
 
diff --git a/src/index/index.rs b/src/index/index.rs
index 81c631cbd..0fc6993fb 100644
--- a/src/index/index.rs
+++ b/src/index/index.rs
@@ -6,6 +6,9 @@ use std::path::PathBuf;
 use std::sync::Arc;
 use std::thread::available_parallelism;
 
+#[cfg(feature = "jitexpr")]
+use jitexpr::compile::ExprCompilationCache;
+
 use super::segment::Segment;
 use super::segment_reader::merge_field_meta_data;
 use super::{FieldMetadata, IndexSettings};
@@ -382,6 +385,8 @@ pub struct Index {
     fast_field_tokenizers: TokenizerManager,
     inventory: SegmentMetaInventory,
     custom_plugins: Vec<Arc<dyn SegmentPlugin>>,
+    #[cfg(feature = "jitexpr")]
+    expr_compilation_cache: ExprCompilationCache,
 }
 
 impl Index {
@@ -502,6 +507,9 @@ impl Index {
             executor: Executor::single_thread(),
             inventory,
             custom_plugins: Vec::new(),
+            // We default to a capacity of 64, but it only allocates if used.
+            #[cfg(feature = "jitexpr")]
+            expr_compilation_cache: ExprCompilationCache::with_capacity(64),
         }
     }
 
@@ -910,8 +918,21 @@ impl Index {
     }
 }
 
+#[cfg(feature = "jitexpr")]
+impl Index {
+    /// Setter for the expression compilation cache.
+    pub fn set_expr_compilation_cache(&mut self, cache: ExprCompilationCache) {
+        self.expr_compilation_cache = cache;
+    }
+
+    /// Accessor for the expression compilation cache.
+    pub fn expr_compilation_cache(&self) -> &ExprCompilationCache {
+        &self.expr_compilation_cache
+    }
+}
+
 impl fmt::Debug for Index {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
         write!(f, "Index({:?})", self.directory)
     }
 }
diff --git a/src/index/inverted_index_plugin.rs b/src/index/inverted_index_plugin.rs
index 68ddf3e48..19273c926 100644
--- a/src/index/inverted_index_plugin.rs
+++ b/src/index/inverted_index_plugin.rs
@@ -17,7 +17,7 @@ use measure_time::debug_time;
 use tokenizer_api::BoxTokenStream;
 
 use crate::directory::{CompositeFile, Directory};
-use crate::docset::{DocSet, TERMINATED};
+use crate::docset::DocSet;
 use crate::error::DataCorruption;
 use crate::fieldnorm::{FieldNormReader, FieldNormReaders, FieldNormsSerializer, FieldNormsWriter};
 use crate::index::{Segment, SegmentComponent, SegmentReader};
@@ -27,7 +27,8 @@ use crate::json_utils::{index_json_value, IndexingPositionsPerPath};
 use crate::plugin::{PluginMergeContext, PluginWriter, PluginWriterContext, SegmentPlugin};
 use crate::postings::{
     compute_table_memory_size, serialize_postings, IndexingContext, IndexingPosition,
-    InvertedIndexSerializer, PerFieldPostingsWriter, Postings, PostingsWriter, SegmentPostings,
+    InvertedIndexSerializer, PerFieldPostingsWriter, Postings, PostingsMerger, PostingsWriter,
+    SegmentPostings,
 };
 use crate::schema::document::{Document, Value};
 use crate::schema::{Field, FieldType, Schema, DATE_TIME_PRECISION_INDEXED};
@@ -595,7 +596,7 @@ fn write_postings_for_field(
     );
 
     let mut segment_postings_containing_the_term: Vec<(usize, SegmentPostings)> = vec![];
-    let mut doc_id_and_positions = vec![];
+    let mut merger = PostingsMerger::new(&merged_doc_id_map);
 
     while merged_terms.advance() {
         segment_postings_containing_the_term.clear();
@@ -645,43 +646,35 @@ fn write_postings_for_field(
 
         field_serializer.new_term(term_bytes, total_doc_freq, has_term_freq)?;
 
-        for (segment_ord, mut segment_postings) in segment_postings_containing_the_term.drain(..) {
-            let old_to_new_doc_id = &merged_doc_id_map[segment_ord];
-
-            let mut doc = segment_postings.doc();
-            while doc != TERMINATED {
-                if let Some(remapped_doc_id) = old_to_new_doc_id[doc as usize] {
+        if doc_id_mapping.is_trivial() {
+            for (segment_ord, mut postings) in segment_postings_containing_the_term.drain(..) {
+                let mapping = &merged_doc_id_map[segment_ord];
+                while let Some(doc) = crate::postings::next_mapped_doc(&mut postings, mapping) {
                     let term_freq = if has_term_freq {
-                        segment_postings.positions(&mut positions_buffer);
-                        segment_postings.term_freq()
+                        postings.positions(&mut positions_buffer);
+                        postings.term_freq()
                     } else {
                         positions_buffer.clear();
-                        0u32
+                        0
                     };
-
-                    if !doc_id_mapping.is_trivial() {
-                        doc_id_and_positions.push((
-                            remapped_doc_id,
-                            term_freq,
-                            positions_buffer.to_vec(),
-                        ));
-                    } else {
-                        let delta_positions = delta_computer.compute_delta(&positions_buffer);
-                        field_serializer.write_doc(remapped_doc_id, term_freq, delta_positions);
-                    }
+                    let delta_positions = delta_computer.compute_delta(&positions_buffer);
+                    field_serializer.write_doc(doc, term_freq, delta_positions);
+                    postings.advance();
                 }
-
-                doc = segment_postings.advance();
             }
-        }
-        if !doc_id_mapping.is_trivial() {
-            doc_id_and_positions.sort_unstable_by_key(|&(doc_id, _, _)| doc_id);
-
-            for (doc_id, term_freq, positions) in &doc_id_and_positions {
-                let delta_positions = delta_computer.compute_delta(positions);
-                field_serializer.write_doc(*doc_id, *term_freq, delta_positions);
+        } else {
+            merger.reset(segment_postings_containing_the_term.drain(..));
+            while merger.advance() {
+                let term_freq = if has_term_freq {
+                    merger.positions(&mut positions_buffer);
+                    merger.term_freq()
+                } else {
+                    positions_buffer.clear();
+                    0
+                };
+                let delta_positions = delta_computer.compute_delta(&positions_buffer);
+                field_serializer.write_doc(merger.doc(), term_freq, delta_positions);
             }
-            doc_id_and_positions.clear();
         }
         field_serializer.close_term()?;
     }
@@ -714,11 +707,12 @@ fn write_postings_merge(
 
 #[cfg(test)]
 mod tests {
-    use super::compute_initial_table_size;
 
-    #[test]
     #[cfg(not(feature = "compare_hash_only"))]
+    #[test]
     fn test_hashmap_size() {
+        use super::compute_initial_table_size;
+
         assert_eq!(compute_initial_table_size(100_000).unwrap(), 1 << 12);
         assert_eq!(compute_initial_table_size(1_000_000).unwrap(), 1 << 15);
         assert_eq!(compute_initial_table_size(15_000_000).unwrap(), 1 << 19);
diff --git a/src/index/segment_reader.rs b/src/index/segment_reader.rs
index 5b6ea0dfc..0aa0c4a96 100644
--- a/src/index/segment_reader.rs
+++ b/src/index/segment_reader.rs
@@ -70,6 +70,11 @@ impl SegmentReader {
         &self.schema
     }
 
+    /// Returns the index this segment belongs to.
+    pub fn index(&self) -> &Index {
+        &self.index
+    }
+
     /// Return the number of documents that have been
     /// deleted in the segment.
     pub fn num_deleted_docs(&self) -> DocId {
diff --git a/src/indexer/merger_sorted_index_test.rs b/src/indexer/merger_sorted_index_test.rs
index 2f230b8df..e717709d4 100644
--- a/src/indexer/merger_sorted_index_test.rs
+++ b/src/indexer/merger_sorted_index_test.rs
@@ -7,15 +7,16 @@ mod tests {
     use crate::collector::TopDocs;
     use crate::fastfield::AliveBitSet;
     use crate::index::Index;
+    use crate::indexer::NoMergePolicy;
     use crate::postings::Postings;
     use crate::query::QueryParser;
     use crate::schema::{
         self, BytesOptions, Facet, FacetOptions, IndexRecordOption, NumericOptions,
-        TextFieldIndexing, TextOptions, Value, FAST, STRING,
+        TextFieldIndexing, TextOptions, Value, FAST, INDEXED, STRING,
     };
     use crate::{
         DocAddress, DocSet, IndexSettings, IndexSortByField, IndexWriter, Order, TantivyDocument,
-        Term,
+        Term, TERMINATED,
     };
 
     fn create_test_index_posting_list_issue(index_settings: Option<IndexSettings>) -> Index {
@@ -920,6 +921,98 @@ mod tests {
         }
     }
 
+    #[test]
+    fn test_merge_sorted_index_postings_with_deletes_and_missing_sort_keys() -> crate::Result<()> {
+        for order in [Order::Asc, Order::Desc] {
+            for record in [
+                IndexRecordOption::Basic,
+                IndexRecordOption::WithFreqs,
+                IndexRecordOption::WithFreqsAndPositions,
+            ] {
+                let mut schema_builder = schema::Schema::builder();
+                let id_field = schema_builder.add_u64_field("id", FAST | INDEXED);
+                let sort_field = schema_builder.add_u64_field("sort", FAST);
+                let text = schema_builder.add_text_field(
+                    "text",
+                    TextOptions::default().set_indexing_options(
+                        TextFieldIndexing::default()
+                            .set_tokenizer("default")
+                            .set_index_option(record),
+                    ),
+                );
+                let index = Index::builder()
+                    .schema(schema_builder.build())
+                    .settings(IndexSettings {
+                        sort_by_field: Some(IndexSortByField {
+                            field: "sort".into(),
+                            order,
+                        }),
+                        ..Default::default()
+                    })
+                    .create_in_ram()?;
+                let mut writer: IndexWriter = index.writer_with_num_threads(1, 15_000_000)?;
+                writer.set_merge_policy(Box::new(NoMergePolicy));
+                for segment in 0..3u64 {
+                    for row in 0..50u64 {
+                        let id = row * 3 + segment;
+                        let mut document =
+                            doc!(id_field=>id, text=>"common anchor ".repeat((id%7+1) as usize));
+                        if id % 5 != 0 {
+                            document.add_u64(sort_field, id % 11);
+                        }
+                        writer.add_document(document)?;
+                    }
+                    writer.commit()?;
+                }
+                for id in (0..150u64).step_by(7) {
+                    writer.delete_term(Term::from_field_u64(id_field, id));
+                }
+                writer.commit()?;
+                let segment_ids = index.searchable_segment_ids()?;
+                assert_eq!(segment_ids.len(), 3);
+                writer.merge(&segment_ids).wait()?;
+
+                let searcher = index.reader()?.searcher();
+                assert_eq!(searcher.segment_readers().len(), 1);
+                let segment = searcher.segment_reader(0);
+                let ids = segment.fast_fields().u64("id")?;
+                let inverted = segment.inverted_index(text)?;
+                for (token, offset) in [("common", 0), ("anchor", 1)] {
+                    let mut postings = inverted
+                        .read_postings(&Term::from_field_text(text, token), record)?
+                        .unwrap();
+                    let mut seen = std::collections::BTreeSet::new();
+                    let mut positions = Vec::new();
+                    let mut previous_doc = None;
+                    while postings.doc() != TERMINATED {
+                        let doc = postings.doc();
+                        if let Some(previous) = previous_doc {
+                            assert!(doc > previous);
+                        }
+                        previous_doc = Some(doc);
+                        let id = ids.first(doc).unwrap();
+                        assert_ne!(id % 7, 0);
+                        assert!(seen.insert(id));
+                        let repetitions = (id % 7 + 1) as u32;
+                        if record != IndexRecordOption::Basic {
+                            assert_eq!(postings.term_freq(), repetitions);
+                        }
+                        if record == IndexRecordOption::WithFreqsAndPositions {
+                            postings.positions(&mut positions);
+                            assert_eq!(
+                                positions,
+                                (0..repetitions).map(|p| 2 * p + offset).collect::<Vec<_>>()
+                            );
+                        }
+                        postings.advance();
+                    }
+                    assert_eq!(seen, (0..150u64).filter(|id| id % 7 != 0).collect());
+                }
+            }
+        }
+        Ok(())
+    }
+
     // #[test]
     // fn test_merge_sorted_index_asc() {
     //     let index = create_test_index(
diff --git a/src/lib.rs b/src/lib.rs
index 53a4fb10f..7537e3cc5 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -220,6 +220,8 @@ use std::fmt;
 
 pub use census::{Inventory, TrackedObject};
 pub use common::{f64_to_u64, i64_to_u64, u64_to_f64, u64_to_i64, HasLen};
+#[cfg(feature = "jitexpr")]
+pub use jitexpr;
 use once_cell::sync::Lazy;
 use serde::{Deserialize, Serialize};
 
diff --git a/src/postings/merger.rs b/src/postings/merger.rs
new file mode 100644
index 000000000..a49254a43
--- /dev/null
+++ b/src/postings/merger.rs
@@ -0,0 +1,305 @@
+//! Merge per-segment postings into increasing mapped document id order.
+
+use std::cmp::Reverse;
+use std::collections::binary_heap::PeekMut;
+use std::collections::BinaryHeap;
+
+use crate::docset::{DocSet, TERMINATED};
+use crate::postings::{Postings, SegmentPostings};
+use crate::DocId;
+
+/// Skip to the next posting whose document is present in `mapping`.
+pub(crate) fn next_mapped_doc(
+    postings: &mut SegmentPostings,
+    mapping: &[Option<DocId>],
+) -> Option<DocId> {
+    while postings.doc() != TERMINATED {
+        if let Some(doc) = mapping[postings.doc() as usize] {
+            return Some(doc);
+        }
+        postings.advance();
+    }
+    None
+}
+
+struct MappedPostings<'a> {
+    postings: SegmentPostings,
+    mapping: &'a [Option<DocId>],
+}
+
+impl MappedPostings<'_> {
+    fn advance(&mut self) -> Option<DocId> {
+        self.postings.advance();
+        next_mapped_doc(&mut self.postings, self.mapping)
+    }
+}
+
+/// Streams postings from several segments in increasing mapped document id.
+///
+/// Create once per field and call [`Self::reset`] for each term.
+pub(crate) struct PostingsMerger<'a> {
+    doc_id_map: &'a [Vec<Option<DocId>>],
+    cursors: Vec<MappedPostings<'a>>,
+    /// Min-heap of `(current mapped doc, index into cursors)`.
+    heap: BinaryHeap<Reverse<(DocId, usize)>>,
+    primed: bool,
+}
+
+impl<'a> PostingsMerger<'a> {
+    /// `doc_id_map[segment_ord][local_doc]` is the mapped doc, or `None` if dropped.
+    pub(crate) fn new(doc_id_map: &'a [Vec<Option<DocId>>]) -> Self {
+        Self {
+            doc_id_map,
+            cursors: Vec::new(),
+            heap: BinaryHeap::new(),
+            primed: false,
+        }
+    }
+
+    /// Start merging a new term from `(segment_ord, postings)` pairs.
+    pub(crate) fn reset(&mut self, segments: impl IntoIterator<Item = (usize, SegmentPostings)>) {
+        self.cursors.clear();
+        self.heap.clear();
+        self.primed = false;
+        let doc_id_map = self.doc_id_map;
+        for (segment_ord, mut postings) in segments {
+            let mapping = &doc_id_map[segment_ord][..];
+            if let Some(doc) = next_mapped_doc(&mut postings, mapping) {
+                self.heap.push(Reverse((doc, self.cursors.len())));
+                self.cursors.push(MappedPostings { postings, mapping });
+            }
+        }
+    }
+
+    /// Advance to the next document. Must return `true` before the accessors are called.
+    pub(crate) fn advance(&mut self) -> bool {
+        if !self.primed {
+            self.primed = true;
+            return !self.heap.is_empty();
+        }
+        let previous = {
+            let Some(mut top) = self.heap.peek_mut() else {
+                return false;
+            };
+            let Reverse((previous, cursor_ord)) = *top;
+            match self.cursors[cursor_ord].advance() {
+                Some(doc) => {
+                    debug_assert!(
+                        doc > previous,
+                        "merge mapping must preserve per-segment order"
+                    );
+                    *top = Reverse((doc, cursor_ord));
+                }
+                None => {
+                    PeekMut::pop(top);
+                }
+            }
+            previous
+        };
+        if let Some(Reverse((next, _))) = self.heap.peek() {
+            debug_assert!(
+                *next > previous,
+                "mapped doc ids must be strictly increasing"
+            );
+        }
+        !self.heap.is_empty()
+    }
+
+    fn current(&self) -> (DocId, usize) {
+        self.heap.peek().expect("advance() returned true").0
+    }
+
+    pub(crate) fn doc(&self) -> DocId {
+        self.current().0
+    }
+
+    pub(crate) fn term_freq(&self) -> u32 {
+        self.cursors[self.current().1].postings.term_freq()
+    }
+
+    pub(crate) fn positions(&mut self, output: &mut Vec<u32>) {
+        let cursor_ord = self.current().1;
+        self.cursors[cursor_ord].postings.positions(output);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::PostingsMerger;
+    use crate::postings::SegmentPostings;
+    use crate::schema::{IndexRecordOption, Schema, TEXT};
+    use crate::{DocId, Index, IndexWriter, Term};
+
+    fn collect(merger: &mut PostingsMerger<'_>) -> Vec<(DocId, u32, Vec<u32>)> {
+        let mut docs = Vec::new();
+        let mut positions = vec![7, 7, 7];
+        while merger.advance() {
+            merger.positions(&mut positions);
+            docs.push((merger.doc(), merger.term_freq(), positions.clone()));
+        }
+        assert!(!merger.advance());
+        docs
+    }
+
+    fn without_positions(docs: &[(DocId, u32)]) -> Vec<(DocId, u32, Vec<u32>)> {
+        docs.iter()
+            .map(|&(doc, tf)| (doc, tf, Vec::new()))
+            .collect()
+    }
+
+    /// Postings with positions for each token, one segment per inner slice.
+    fn postings_with_positions(
+        segments: &[&[&str]],
+        tokens: &[&str],
+    ) -> crate::Result<Vec<Vec<(usize, SegmentPostings)>>> {
+        let mut schema_builder = Schema::builder();
+        let text = schema_builder.add_text_field("text", TEXT);
+        let schema = schema_builder.build();
+        let mut readers = Vec::new();
+        for docs in segments {
+            let index = Index::create_in_ram(schema.clone());
+            let mut writer: IndexWriter = index.writer_for_tests()?;
+            for body in *docs {
+                writer.add_document(doc!(text => *body))?;
+            }
+            writer.commit()?;
+            let searcher = index.reader()?.searcher();
+            assert_eq!(searcher.segment_readers().len(), 1);
+            readers.push(searcher.segment_reader(0).inverted_index(text)?);
+        }
+        let mut terms = Vec::new();
+        for token in tokens {
+            let term = Term::from_field_text(text, token);
+            let mut postings = Vec::new();
+            for (segment_ord, reader) in readers.iter().enumerate() {
+                if let Some(segment_postings) =
+                    reader.read_postings(&term, IndexRecordOption::WithFreqsAndPositions)?
+                {
+                    postings.push((segment_ord, segment_postings));
+                }
+            }
+            terms.push(postings);
+        }
+        Ok(terms)
+    }
+
+    #[test]
+    fn test_merges_segments_skipping_deletes() {
+        // Segment 2 is fully deleted and segment 3 has no postings.
+        let doc_id_map = vec![
+            vec![Some(1), None, Some(3)],
+            vec![Some(0), Some(2)],
+            vec![None, None],
+            Vec::new(),
+        ];
+        let segments = [
+            (
+                0,
+                SegmentPostings::create_from_docs_and_tfs(&[(0, 1), (1, 2), (2, 3)], None),
+            ),
+            (
+                1,
+                SegmentPostings::create_from_docs_and_tfs(&[(0, 4), (1, 5)], None),
+            ),
+            (
+                2,
+                SegmentPostings::create_from_docs_and_tfs(&[(0, 9), (1, 9)], None),
+            ),
+            (3, SegmentPostings::empty()),
+        ];
+        let mut merger = PostingsMerger::new(&doc_id_map);
+        merger.reset(segments);
+        assert_eq!(
+            collect(&mut merger),
+            without_positions(&[(0, 4), (1, 1), (2, 5), (3, 3)])
+        );
+    }
+
+    #[test]
+    fn test_empty_term_yields_nothing() {
+        let doc_id_map = vec![vec![Some(0)]];
+        let mut merger = PostingsMerger::new(&doc_id_map);
+        merger.reset([(0, SegmentPostings::empty())]);
+        assert_eq!(collect(&mut merger), Vec::new());
+    }
+
+    #[test]
+    fn test_crosses_postings_block_boundary() {
+        const N: u32 = 200;
+        let seg0: Vec<(u32, u32)> = (0..N).map(|doc| (doc, doc + 1)).collect();
+        let seg1: Vec<(u32, u32)> = (0..N).map(|doc| (doc, 1_000 + doc)).collect();
+        let map0: Vec<Option<DocId>> = (0..N)
+            .map(|doc| if doc % 10 == 0 { None } else { Some(doc * 2) })
+            .collect();
+        let map1: Vec<Option<DocId>> = (0..N).map(|doc| Some(doc * 2 + 1)).collect();
+        let doc_id_map = vec![map0, map1];
+        let mut merger = PostingsMerger::new(&doc_id_map);
+        merger.reset([
+            (0, SegmentPostings::create_from_docs_and_tfs(&seg0, None)),
+            (1, SegmentPostings::create_from_docs_and_tfs(&seg1, None)),
+        ]);
+
+        let mut expected = Vec::new();
+        for doc in 0..N {
+            if doc % 10 != 0 {
+                expected.push((doc * 2, doc + 1));
+            }
+            expected.push((doc * 2 + 1, 1_000 + doc));
+        }
+        expected.sort_unstable();
+        assert_eq!(collect(&mut merger), without_positions(&expected));
+    }
+
+    #[test]
+    fn test_positions_follow_their_document() -> crate::Result<()> {
+        let segments: [&[&str]; 2] = [&["a b a", "b", "b a b a a"], &["a", "b b a", "a b", "b"]];
+        let doc_id_map = vec![
+            vec![Some(1), None, Some(4)],
+            vec![Some(0), Some(2), None, Some(3)],
+        ];
+        let mut terms = postings_with_positions(&segments, &["a", "b"])?.into_iter();
+        let mut merger = PostingsMerger::new(&doc_id_map);
+
+        merger.reset(terms.next().unwrap());
+        assert_eq!(
+            collect(&mut merger),
+            vec![
+                (0, 1, vec![0]),
+                (1, 2, vec![0, 2]),
+                (2, 1, vec![2]),
+                (4, 3, vec![1, 3, 4]),
+            ]
+        );
+
+        merger.reset(terms.next().unwrap());
+        assert_eq!(
+            collect(&mut merger),
+            vec![
+                (1, 1, vec![1]),
+                (2, 2, vec![0, 1]),
+                (3, 1, vec![0]),
+                (4, 2, vec![0, 2]),
+            ]
+        );
+        Ok(())
+    }
+
+    #[test]
+    fn test_reset_discards_unfinished_term() -> crate::Result<()> {
+        let segments: [&[&str]; 2] = [&["a b", "a"], &["b a", "a b", "a"]];
+        let doc_id_map = vec![vec![Some(0), Some(2)], vec![Some(1), Some(3), Some(4)]];
+        let mut terms = postings_with_positions(&segments, &["a", "b"])?.into_iter();
+        let mut merger = PostingsMerger::new(&doc_id_map);
+
+        merger.reset(terms.next().unwrap());
+        assert!(merger.advance());
+        assert_eq!(merger.doc(), 0);
+
+        merger.reset(terms.next().unwrap());
+        assert_eq!(
+            collect(&mut merger),
+            vec![(0, 1, vec![1]), (1, 1, vec![0]), (3, 1, vec![1])]
+        );
+        Ok(())
+    }
+}
diff --git a/src/postings/mod.rs b/src/postings/mod.rs
index ea512230c..61f83e62f 100644
--- a/src/postings/mod.rs
+++ b/src/postings/mod.rs
@@ -9,6 +9,7 @@ pub(crate) mod compression;
 mod indexing_context;
 mod json_postings_writer;
 mod loaded_postings;
+mod merger;
 mod per_field_postings_writer;
 mod postings;
 mod postings_writer;
@@ -20,6 +21,7 @@ mod skip;
 mod term_info;
 
 pub(crate) use loaded_postings::LoadedPostings;
+pub(crate) use merger::{next_mapped_doc, PostingsMerger};
 pub(crate) use stacker::compute_table_memory_size;
 
 pub use self::block_segment_postings::BlockSegmentPostings;
diff --git a/src/postings/serializer.rs b/src/postings/serializer.rs
index f44d14d93..4fbd42620 100644
--- a/src/postings/serializer.rs
+++ b/src/postings/serializer.rs
@@ -249,7 +249,6 @@ impl<'a, W: Write> FieldSerializer<'a, W> {
         if let Some(positions_serializer) = self.positions_serializer_opt {
             positions_serializer.close()?;
         }
-        self.postings_write.flush()?;
         self.term_dictionary_builder.finish()?;
         Ok(())
     }
diff --git a/src/query/all_query.rs b/src/query/all_query.rs
index 5431a3a1b..e4749ac24 100644
--- a/src/query/all_query.rs
+++ b/src/query/all_query.rs
@@ -47,7 +47,8 @@ pub struct AllScorer {
 impl AllScorer {
     /// Creates a new AllScorer with `max_doc` docs.
     pub fn new(max_doc: DocId) -> AllScorer {
-        AllScorer { doc: 0u32, max_doc }
+        let doc = if max_doc == 0u32 { TERMINATED } else { 0 };
+        AllScorer { doc, max_doc }
     }
 }
 
diff --git a/src/query/doc_predicate_query/function_predicate.rs b/src/query/doc_predicate_query/function_predicate.rs
index ac7192580..acd5fd206 100644
--- a/src/query/doc_predicate_query/function_predicate.rs
+++ b/src/query/doc_predicate_query/function_predicate.rs
@@ -1,5 +1,7 @@
 use super::{DocPredicate, SegmentDocPredicate};
 use crate::index::SegmentReader;
+use crate::query::doc_predicate_query::ConstOrVariableSegmentPredicate;
+use crate::query::AllScorer;
 use crate::DocId;
 
 /// Blanket [`SegmentDocPredicate`] implementation for any per-document
@@ -48,8 +50,15 @@ where
 {
     type SegmentDocPredicate = SegmentF;
 
-    fn doc_predicate(&self, segment_reader: &SegmentReader) -> crate::Result<SegmentF> {
-        (self.segment_predicate_factory)(segment_reader)
+    fn doc_predicate(
+        &self,
+        segment_reader: &SegmentReader,
+    ) -> crate::Result<ConstOrVariableSegmentPredicate<SegmentF>> {
+        let predicate = (self.segment_predicate_factory)(segment_reader)?;
+        Ok(ConstOrVariableSegmentPredicate::Variable {
+            predicate,
+            necessary_condition: Box::new(AllScorer::new(segment_reader.max_doc())),
+        })
     }
 }
 
diff --git a/src/query/doc_predicate_query/jitexpr_predicate.rs b/src/query/doc_predicate_query/jitexpr_predicate.rs
new file mode 100644
index 000000000..9698ac3fd
--- /dev/null
+++ b/src/query/doc_predicate_query/jitexpr_predicate.rs
@@ -0,0 +1,836 @@
+use std::collections::HashMap;
+use std::io;
+
+use columnar::{ColumnIndex, ColumnType, DynamicColumn, StrColumn};
+use jitexpr::ast::{
+    infer_types_with_target, required_presence_for_true, InferredTypeSet, TypeError, UntypedExpr,
+    VariablePresenceCondition,
+};
+use jitexpr::compile::{CompiledFnCtx, StringArena};
+use jitexpr::types::{VarType, VariableValue};
+
+use super::{DocPredicate, SegmentDocPredicate};
+use crate::index::SegmentReader;
+use crate::query::doc_predicate_query::ConstOrVariableSegmentPredicate;
+use crate::query::exist_query::{ExistsColumnIndex, ExistsDocSet};
+use crate::query::union::SimpleUnion;
+use crate::query::{AllScorer, EmptyScorer, Intersection};
+use crate::{DocId, DocSet, TantivyError, TERMINATED};
+
+/// A [`DocPredicate`] that evaluates a boolean JIT expression against fast fields.
+///
+/// Requires the `jitexpr` feature. Variable names are resolved as fast-field names
+/// for each segment, supporting boolean, numeric, and string columns. Missing or
+/// incompatible columns are left unbound, so the compiler treats them as `None`.
+///
+/// For bound columns, multivalued documents contribute their first value, and a
+/// document missing any input will still be evaluated with null in place of the input.
+///
+/// We "fast path" cases where we detect the expression will always evaluate to true or false
+/// on a segment. (e.g. if all variable columns are missing).
+///
+/// Only a present `true` result matches.
+///
+/// Documents missing the fields required for the expression to be `true` are skipped without
+/// being evaluated. For instance, `(EQ (ADD price 1u64) 10u64)` is only evaluated on the
+/// documents having a `price` value.
+#[derive(Clone, Debug)]
+pub struct JitExprPredicate {
+    expression: UntypedExpr,
+    inferred_inputs: Vec<(String, InferredTypeSet)>,
+    // A necessary condition, on the presence of the variables, for the expression to be `true`.
+    required_presence: VariablePresenceCondition,
+}
+
+impl JitExprPredicate {
+    /// Creates a predicate after inferring its inputs and requiring a boolean result.
+    pub fn new(expression: UntypedExpr) -> Result<Self, TypeError> {
+        let inferred_types: HashMap<&str, InferredTypeSet> =
+            infer_types_with_target(&expression, InferredTypeSet::BOOLEAN)?;
+        let inferred_inputs: Vec<(String, InferredTypeSet)> = inferred_types
+            .into_iter()
+            .map(|(name, types)| (name.to_string(), types))
+            .collect();
+        let required_presence = required_presence_for_true(&expression);
+        Ok(Self {
+            expression,
+            inferred_inputs,
+            required_presence,
+        })
+    }
+
+    /// Returns the expression evaluated by this predicate.
+    pub fn expression(&self) -> &UntypedExpr {
+        &self.expression
+    }
+}
+
+impl DocPredicate for JitExprPredicate {
+    type SegmentDocPredicate = JitExprEvalState;
+
+    fn doc_predicate(
+        &self,
+        segment_reader: &SegmentReader,
+    ) -> crate::Result<ConstOrVariableSegmentPredicate<JitExprEvalState>> {
+        let mut variable_types = HashMap::with_capacity(self.inferred_inputs.len());
+        let mut opened_columns: HashMap<&str, DynamicColumn> =
+            HashMap::with_capacity(self.inferred_inputs.len());
+
+        // We pick a single column for each variable name. NOTE this CAN yield to unexpected results
+        // for some expression (e.g. (IS_NULL "mycol")).
+        // For instance, a document could be matching in one segment, and not matching if it
+        // was in another segment, just because the presence of column with the same name
+        // and different type could interfere.
+        for (name, accepted_types) in &self.inferred_inputs {
+            let Some(column) = open_input_column(segment_reader, name, *accepted_types)? else {
+                // If we do not have a valid column for that expression, we do not
+                // fill the HashMap at all.
+                //
+                // The compiler will replace the expression and make it behave like the null
+                // literal.
+                continue;
+            };
+            let Some(var_type) = var_type_for_column_type(column.column_type()) else {
+                continue;
+            };
+            variable_types.insert(name.as_str(), var_type);
+            opened_columns.insert(name.as_str(), column);
+        }
+
+        // The variables are bound to the columns opened above, and only to them: the presence
+        // of a variable is the presence of a value in its column.
+        let necessary_condition: Box<dyn DocSet> = build_necessary_condition_docset(
+            &self.required_presence,
+            &opened_columns,
+            segment_reader.max_doc(),
+        );
+        if necessary_condition.doc() == TERMINATED {
+            return Ok(ConstOrVariableSegmentPredicate::Const(false));
+        }
+
+        let compiled_fn = segment_reader
+            .index()
+            .expr_compilation_cache()
+            .compile(&self.expression, &variable_types)
+            .map_err(|compilation_err| {
+                TantivyError::InvalidArgument(format!(
+                    "the expression compilation failed {:?}. error: {compilation_err}",
+                    self.expression
+                ))
+            })?;
+
+        // If the function is by nature const, or if all of its variable are known to be null
+        // (because we don't have such columns), then we eval the value only once optimize
+        //
+        // TODO optimize further when the columns have a single value (full + min/max value)
+        if variable_types.is_empty() || compiled_fn.inputs().is_empty() {
+            // We have no variables!
+            // This means eventual inputs are not. Let's return a const predicate.
+            let inputs: Vec<VariableValue> =
+                std::iter::repeat_n(VariableValue::none(), compiled_fn.inputs().len()).collect();
+            let mut string_arena = StringArena::default();
+            let result = unsafe { compiled_fn.call(&inputs[..], &mut string_arena) };
+            let const_bool = unsafe { result.as_bool() }.unwrap_or(false);
+            return Ok(ConstOrVariableSegmentPredicate::Const(const_bool));
+        }
+
+        // We ended up with an expression that could not resolve to anything apparently.
+        if compiled_fn.result_type() == VarType::None {
+            return Ok(ConstOrVariableSegmentPredicate::Const(false));
+        }
+
+        if compiled_fn.result_type() != VarType::Bool {
+            // This should never happen: we passed a target inferred type of Bool,
+            // so we should have either Bool or None.
+            return Err(TantivyError::InvalidArgument(format!(
+                "the expression is not a predicate {}",
+                self.expression
+            )));
+        }
+
+        // The compiler owns the definitive ABI order. Do not rely on inference
+        // or HashMap iteration order when building the argument slots.
+        let mut columns_opt = Vec::with_capacity(compiled_fn.inputs().len());
+        for input in compiled_fn.inputs() {
+            let column_opt: Option<DynamicColumn> =
+                opened_columns.remove(input.variable_name.as_ref());
+            if let Some(column) = column_opt {
+                if var_type_for_column_type(column.column_type()) != Some(input.r#type) {
+                    return Err(TantivyError::InternalError(format!(
+                        "compiled input `{}` expects {:?}, but its column has type {}",
+                        input.variable_name,
+                        input.r#type,
+                        column.column_type()
+                    )));
+                }
+                columns_opt.push(Some(column));
+            } else {
+                columns_opt.push(None);
+            }
+        }
+        // There is one reusable buffer per string column, in ABI order.
+        let num_string_inputs = columns_opt
+            .iter()
+            .filter(|column_opt| matches!(column_opt, Some(DynamicColumn::Str(_))))
+            .count();
+        let num_inputs = columns_opt.len();
+        let predicate = JitExprEvalState {
+            compiled: compiled_fn.context(),
+            columns_opt,
+            string_inputs: vec![String::new(); num_string_inputs],
+            input_values: Vec::with_capacity(num_inputs),
+        };
+        Ok(ConstOrVariableSegmentPredicate::Variable {
+            predicate,
+            necessary_condition,
+        })
+    }
+}
+
+/// Builds a docset off the variable presence condition.
+///
+/// Variables are bound to the columns of `columns`. A variable missing from `columns` is null
+/// for all documents.
+///
+/// The returned `DocSet` is positioned on its first document. It is `TERMINATED` if and only if
+/// no document of the segment satisfies the condition.
+fn build_necessary_condition_docset(
+    condition: &VariablePresenceCondition,
+    columns: &HashMap<&str, DynamicColumn>,
+    max_doc: DocId,
+) -> Box<dyn DocSet> {
+    match condition {
+        VariablePresenceCondition::Always => Box::new(AllScorer::new(max_doc)),
+        VariablePresenceCondition::Never => Box::new(EmptyScorer),
+        VariablePresenceCondition::Present(variable_name) => {
+            let Some(column) = columns.get(variable_name.as_ref()) else {
+                return Box::new(EmptyScorer);
+            };
+            let exists_column_index = match column.column_index() {
+                ColumnIndex::Empty { .. } => return Box::new(EmptyScorer),
+                ColumnIndex::Full => return Box::new(AllScorer::new(max_doc)),
+                ColumnIndex::Optional(optional_index) => {
+                    ExistsColumnIndex::Optional(optional_index.clone())
+                }
+                ColumnIndex::Multivalued(multivalued_index) => {
+                    ExistsColumnIndex::Multivalued(multivalued_index.clone())
+                }
+            };
+            Box::new(ExistsDocSet::new(exists_column_index))
+        }
+        VariablePresenceCondition::All(conditions) => {
+            let mut doc_sets: Vec<Box<dyn DocSet>> = conditions
+                .iter()
+                .map(|condition| build_necessary_condition_docset(condition, columns, max_doc))
+                .collect();
+            match doc_sets.len() {
+                0 => Box::new(AllScorer::new(max_doc)),
+                1 => doc_sets.pop().unwrap(),
+                _ => Box::new(Intersection::new(doc_sets, max_doc)),
+            }
+        }
+        VariablePresenceCondition::Any(conditions) => {
+            let doc_sets: Vec<Box<dyn DocSet>> = conditions
+                .iter()
+                .map(|condition| build_necessary_condition_docset(condition, columns, max_doc))
+                .collect();
+            Box::new(SimpleUnion::build(doc_sets))
+        }
+    }
+}
+
+fn open_input_column(
+    reader: &SegmentReader,
+    name: &str,
+    accepted_types: InferredTypeSet,
+) -> io::Result<Option<DynamicColumn>> {
+    let Ok(column_handles) = reader.fast_fields().dynamic_column_handles(name) else {
+        // If the call to dynamic_column_handles fails (for instance because the column is not a
+        // fast field) we choose to act as if the column was absent.
+        return Ok(None);
+    };
+    for handle in column_handles {
+        // We return the first column that could be accepted
+        let Some(var_type) = var_type_for_column_type(handle.column_type()) else {
+            continue;
+        };
+        if accepted_types.contains(var_type) {
+            return Ok(Some(handle.open()?));
+        }
+    }
+    Ok(None)
+}
+
+fn var_type_for_column_type(column_type: ColumnType) -> Option<VarType> {
+    match column_type {
+        ColumnType::Bool => Some(VarType::Bool),
+        ColumnType::I64 => Some(VarType::I64),
+        ColumnType::U64 => Some(VarType::U64),
+        ColumnType::F64 => Some(VarType::F64),
+        ColumnType::Str => Some(VarType::Str),
+        ColumnType::Bytes | ColumnType::IpAddr | ColumnType::DateTime => None,
+    }
+}
+
+/// The [`SegmentDocPredicate`] produced by [`JitExprPredicate`] for one segment.
+pub struct JitExprEvalState {
+    compiled: CompiledFnCtx,
+    columns_opt: Vec<Option<DynamicColumn>>,
+    // One reusable buffer per string column.
+    string_inputs: Vec<String>,
+    // Reusable argument slots.
+    //
+    // Hidden contract: this vector is always empty between evaluations, so the
+    // `'static` lifetime is a placeholder for an unused element type rather
+    // than a claim about any stored string. Only its capacity carries over,
+    // which is what makes restoring the `'static` type after an evaluation
+    // sound. `eval` is responsible for upholding this on every return path.
+    input_values: Vec<VariableValue<'static>>,
+}
+
+/// A wrapper to make sure the variable value buffer is cleared even if the evaluation
+/// panicked.
+struct ClearOnDrop<'a>(&'a mut Vec<VariableValue<'a>>);
+
+impl<'a> ClearOnDrop<'a> {
+    fn wrap(input_values: &'a mut Vec<VariableValue<'static>>) -> Self {
+        debug_assert!(input_values.is_empty());
+        // Input_values is just a buffer we share to avoid allocations
+        let lower_lifetime_input_values: &mut Vec<VariableValue<'_>> =
+            unsafe { std::mem::transmute(input_values) };
+        ClearOnDrop(lower_lifetime_input_values)
+    }
+}
+
+impl<'a> Drop for ClearOnDrop<'a> {
+    fn drop(&mut self) {
+        self.0.clear();
+    }
+}
+
+impl SegmentDocPredicate for JitExprEvalState {
+    fn eval(&mut self, doc_id: DocId) -> bool {
+        // Input_values is just a buffer we share to avoid allocations
+        let inputs_vec = ClearOnDrop::wrap(&mut self.input_values);
+
+        fill_input_values(
+            &self.columns_opt,
+            &mut self.string_inputs,
+            inputs_vec.0,
+            doc_id,
+        );
+
+        // SAFETY: Columns follow compiled.inputs() and their types were checked
+        // during setup. Each slot uses the matching union arm. String buffers
+        // remain borrowed, and cannot be mutated, until this call finishes.
+        let eval_result: Option<bool> = unsafe { self.compiled.call(inputs_vec.0).as_bool() };
+
+        eval_result == Some(true)
+    }
+}
+
+fn fill_input_values<'buffer>(
+    columns: &[Option<DynamicColumn>],
+    string_inputs: &'buffer mut [String],
+    input_values: &mut Vec<VariableValue<'buffer>>,
+    doc_id: DocId,
+) {
+    debug_assert!(input_values.is_empty());
+    let mut string_inputs = string_inputs.iter_mut();
+    for column_opt in columns {
+        let Some(column) = column_opt else {
+            // The full column is absent. We treat it as None.
+            input_values.push(VariableValue::none());
+            continue;
+        };
+        let input: Option<VariableValue> = match column {
+            DynamicColumn::Bool(column) => column.first(doc_id).map(VariableValue::from),
+            DynamicColumn::I64(column) => column.first(doc_id).map(VariableValue::from),
+            DynamicColumn::U64(column) => column.first(doc_id).map(VariableValue::from),
+            DynamicColumn::F64(column) => column.first(doc_id).map(VariableValue::from),
+            DynamicColumn::Str(column) => {
+                let string_input = string_inputs
+                    .next()
+                    .expect("every string column has a string input buffer");
+                load_str_input(column, doc_id, string_input).map(VariableValue::from)
+            }
+            DynamicColumn::Bytes(_) | DynamicColumn::IpAddr(_) | DynamicColumn::DateTime(_) => {
+                unreachable!("unsupported columns are filtered before compilation")
+            }
+        };
+        // If the value is someone absent, we set the input to none/null.
+        input_values.push(input.unwrap_or(VariableValue::none()));
+    }
+}
+
+/// Loads the first value of a string column for `doc_id` into `buffer`.
+///
+/// Missing values return None.
+///
+/// This function may panic if the dictionary is corrupted or if the column
+/// contains term ords that do not exist in the dictionary.
+fn load_str_input<'buffer>(
+    column: &StrColumn,
+    doc_id: DocId,
+    buffer: &'buffer mut String,
+) -> Option<&'buffer str> {
+    buffer.clear();
+    let term_ord = column.ords().first(doc_id)?;
+    // SegmentDocPredicate::eval cannot return I/O errors; an unreadable
+    // dictionary therefore panics.
+    // TODO this is terribly inefficient: we need at least some caching.
+    let found = column
+        .ord_to_str(term_ord, buffer)
+        .expect("fast-field string dictionary is corrupted");
+    assert!(found);
+    Some(buffer.as_str())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::collector::{Count, DocSetCollector};
+    use crate::query::doc_predicate_query::DocPredicateQuery;
+    use crate::query::{EnableScoring, Query};
+    use crate::schema::{Schema, FAST, INDEXED, STORED, STRING};
+    use crate::{Index, TantivyDocument, Term};
+
+    fn create_index() -> Index {
+        let mut schema_builder = Schema::builder();
+        let number = schema_builder.add_u64_field("number", FAST);
+        let flag = schema_builder.add_bool_field("flag", FAST);
+        let _notfast = schema_builder.add_bool_field("notfast", STORED);
+        let label = schema_builder.add_text_field("label", STRING | FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut writer = index.writer_for_tests().unwrap();
+        writer
+            .add_document(doc!(number => 1u64, flag => true, label => "one"))
+            .unwrap();
+        writer
+            .add_document(doc!(number => 2u64, flag => false, label => "two"))
+            .unwrap();
+        writer
+            .add_document(doc!(number => 3u64, flag => true, label => "three"))
+            .unwrap();
+        writer.add_document(doc!(number => 4u64)).unwrap();
+        writer.commit().unwrap();
+        index
+    }
+
+    fn query(expression: &str) -> DocPredicateQuery {
+        JitExprPredicate::new(jitexpr::ast::deserialize(expression).unwrap())
+            .unwrap()
+            .into()
+    }
+
+    #[test]
+    fn test_constructor_requires_boolean_expression() {
+        let expression = jitexpr::ast::deserialize("(ADD number 1u64)").unwrap();
+        assert!(JitExprPredicate::new(expression).is_err());
+    }
+
+    #[test]
+    fn test_simple() {
+        let index = create_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(
+            searcher.search(&query("(EQ number 2i64)"), &Count).unwrap(),
+            1
+        );
+        assert_eq!(
+            searcher
+                .search(&query("(EQ (ADD number 1u64) 3u64)"), &Count)
+                .unwrap(),
+            1
+        );
+    }
+
+    #[test]
+    fn test_simple_string_ref_predicate() {
+        let index = create_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(
+            searcher
+                .search(
+                    &query(r#"(EQ (REGEXP_EXTRACT label "(.).*" 1u64) "o")"#),
+                    &Count
+                )
+                .unwrap(),
+            1
+        );
+    }
+
+    #[test]
+    fn test_simple_built_string_predicate() {
+        let index = create_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(
+            searcher
+                .search(&query(r#"(EQ (UPPER label) "TWO")"#), &Count)
+                .unwrap(),
+            1
+        );
+    }
+
+    #[test]
+    fn test_simple_missing_field() {
+        let index = create_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(
+            searcher
+                .search(&query(r#"(EQ missing_field true)"#), &Count)
+                .unwrap(),
+            0
+        );
+    }
+
+    #[test]
+    fn test_simple_missing_field_is_not_null() {
+        let index = create_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(
+            searcher
+                .search(&query(r#"(IS_NOT_NULL missing_field)"#), &Count)
+                .unwrap(),
+            0
+        );
+    }
+
+    #[test]
+    fn test_simple_missing_field_is_null() {
+        let index = create_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(
+            searcher
+                .search(&query(r#"(IS_NULL missing_field)"#), &Count)
+                .unwrap(),
+            4
+        );
+    }
+
+    #[test]
+    fn test_simple_notfast() {
+        let index = create_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(
+            searcher
+                .search(&query(r#"(EQ notfast true)"#), &Count)
+                .unwrap(),
+            0
+        );
+    }
+
+    #[test]
+    fn test_boolean_and_string_inputs_follow_compiled_order() {
+        let index = create_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(searcher.search(&query("flag"), &Count).unwrap(), 2);
+        assert_eq!(
+            searcher
+                .search(&query(r#"(EQ label "three")"#), &Count)
+                .unwrap(),
+            1
+        );
+        // Inference sorts names, but the ABI follows expression order: label, flag.
+        assert_eq!(
+            searcher
+                .search(&query(r#"(EQ (EQ label "three") flag)"#), &Count)
+                .unwrap(),
+            2
+        );
+    }
+
+    #[test]
+    fn test_constant_predicates() {
+        let index = create_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(searcher.search(&query("true"), &Count).unwrap(), 4);
+        assert_eq!(searcher.search(&query("false"), &Count).unwrap(), 0);
+    }
+
+    #[test]
+    fn test_missing_and_incompatible_columns() {
+        let index = create_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(searcher.search(&query("missing"), &Count).unwrap(), 0);
+
+        // A column missing from the segment is compiled as None.
+        assert_eq!(
+            searcher
+                .search(&query("(IS_NULL missing)"), &Count)
+                .unwrap(),
+            4
+        );
+        // A column missing from the segment is compiled as None.
+        assert_eq!(
+            searcher
+                .search(&query("(IS_NOT_NULL missing)"), &Count)
+                .unwrap(),
+            0
+        );
+        assert_eq!(
+            searcher.search(&query("(IS_NULL flag)"), &Count).unwrap(),
+            1
+        );
+        assert_eq!(
+            searcher
+                .search(&query("(IS_NOT_NULL flag)"), &Count)
+                .unwrap(),
+            3
+        );
+        assert_eq!(
+            searcher
+                .search(&query("(EQ (ADD label 1i64) 2i64)"), &Count)
+                .unwrap(),
+            0
+        );
+        assert_eq!(
+            searcher
+                .search(&query("(IS_NULL (ADD label 1i64))"), &Count)
+                .unwrap(),
+            4
+        );
+    }
+
+    #[test]
+    fn test_signed_and_float_columns() {
+        let mut schema_builder = Schema::builder();
+        let signed = schema_builder.add_i64_field("signed", FAST);
+        let float = schema_builder.add_f64_field("float", FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut writer = index.writer_for_tests().unwrap();
+        writer
+            .add_document(doc!(signed => -2i64, float => 1.5f64))
+            .unwrap();
+        writer
+            .add_document(doc!(signed => 3i64, float => 2.5f64))
+            .unwrap();
+        writer.add_document(doc!()).unwrap();
+        writer.commit().unwrap();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(
+            searcher
+                .search(&query("(EQ signed -2i64)"), &Count)
+                .unwrap(),
+            1
+        );
+        assert_eq!(
+            searcher
+                .search(&query("(EQ signed -2f64)"), &Count)
+                .unwrap(),
+            1
+        );
+        assert_eq!(
+            searcher
+                .search(&query("(EQ float 1.5f64)"), &Count)
+                .unwrap(),
+            1
+        );
+        assert_eq!(
+            searcher.search(&query("(EQ float 1i64)"), &Count).unwrap(),
+            0
+        );
+    }
+
+    #[test]
+    fn test_multivalued_columns_use_first_value() {
+        let mut schema_builder = Schema::builder();
+        let number = schema_builder.add_u64_field("number", FAST);
+        let label = schema_builder.add_text_field("label", STRING | FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut writer = index.writer_for_tests().unwrap();
+        writer
+            .add_document(doc!(number => 1u64, number => 2u64,
+                                 label => "first", label => "second"))
+            .unwrap();
+        writer
+            .add_document(doc!(number => 2u64, number => 1u64,
+                                 label => "second", label => "first"))
+            .unwrap();
+        writer.add_document(doc!()).unwrap();
+        writer.commit().unwrap();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(
+            searcher.search(&query("(EQ number 1u64)"), &Count).unwrap(),
+            1
+        );
+        assert_eq!(
+            searcher
+                .search(&query(r#"(EQ label "first")"#), &Count)
+                .unwrap(),
+            1
+        );
+    }
+
+    /// Two segments with sparse, multivalued, and segment-dependent columns, and deleted docs.
+    ///
+    /// `label` only has values in the second segment.
+    fn create_sparse_index() -> Index {
+        let mut schema_builder = Schema::builder();
+        let id = schema_builder.add_u64_field("id", FAST | INDEXED);
+        let number = schema_builder.add_u64_field("number", FAST);
+        let score = schema_builder.add_i64_field("score", FAST);
+        let flag = schema_builder.add_bool_field("flag", FAST);
+        let label = schema_builder.add_text_field("label", STRING | FAST);
+        let tags = schema_builder.add_text_field("tags", STRING | FAST);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut writer = index.writer_for_tests().unwrap();
+        for segment_ord in 0..2u64 {
+            for i in 0..300u64 {
+                let mut doc = TantivyDocument::default();
+                doc.add_u64(id, segment_ord * 1000 + i);
+                if i % 3 == 0 {
+                    doc.add_u64(number, i);
+                }
+                if i % 5 == 0 {
+                    doc.add_i64(score, (i % 7) as i64 - 3);
+                }
+                if i % 2 == 0 {
+                    doc.add_bool(flag, i % 4 == 0);
+                }
+                if segment_ord == 1 && i % 4 == 0 {
+                    doc.add_text(label, ["a", "b", "ab"][(i % 3) as usize]);
+                }
+                if i % 6 == 0 {
+                    doc.add_text(tags, "x");
+                    doc.add_text(tags, "y");
+                } else if i % 6 == 1 {
+                    doc.add_text(tags, "y");
+                }
+                writer.add_document(doc).unwrap();
+            }
+            writer.commit().unwrap();
+        }
+        writer.delete_term(Term::from_field_u64(id, 30));
+        writer.delete_term(Term::from_field_u64(id, 1060));
+        writer.commit().unwrap();
+        index
+    }
+
+    /// Returns a query matching the same documents as `query(expression)`, but requiring the
+    /// presence of no field, so that all documents are evaluated.
+    ///
+    /// `(NOT true)` is a present `false`: the disjunction is `true` if and only if `expression` is.
+    /// As `NOT` requires the presence of no field, neither does the disjunction.
+    fn query_without_required_presence(expression: &str) -> DocPredicateQuery {
+        query(&format!("(OR {expression} (NOT true))"))
+    }
+
+    #[test]
+    fn test_required_presence_does_not_change_results() {
+        let index = create_sparse_index();
+        let searcher = index.reader().unwrap().searcher();
+        assert_eq!(searcher.segment_readers().len(), 2);
+        let expressions = [
+            "true",
+            "false",
+            "flag",
+            "(EQ number 33u64)",
+            "(EQ number 30u64)",
+            "(IS_NOT_NULL number)",
+            "(IS_NULL score)",
+            "(NOT (EQ number 3u64))",
+            "(NEQ score 1i64)",
+            "(GT (ADD number score) 10i64)",
+            "(OR (EQ number 3u64) (EQ score 1i64))",
+            "(AND flag (IS_NOT_NULL label))",
+            r#"(EQ label "a")"#,
+            r#"(OR (EQ label "b") flag)"#,
+            r#"(REGEXP_LIKE label "b")"#,
+            r#"(EQ (UPPER tags) "X")"#,
+            "(IF flag (GT number 100u64) (LT score 0i64))",
+            "(IS_NOT_NULL (IF flag number score))",
+            "(AND (EQ missing 1i64) flag)",
+            "(OR (IS_NOT_NULL missing) (EQ score 2i64))",
+            "(OR (IS_NULL missing) (EQ score 2i64))",
+        ];
+        for expression in expressions {
+            let expected = searcher
+                .search(
+                    &query_without_required_presence(expression),
+                    &DocSetCollector,
+                )
+                .unwrap();
+            let accelerated = searcher
+                .search(&query(expression), &DocSetCollector)
+                .unwrap();
+            assert_eq!(accelerated, expected, "{expression}");
+        }
+        // Sanity checks: the index does exercise the predicates.
+        let count = |expression: &str| searcher.search(&query(expression), &Count).unwrap();
+        assert_eq!(count("(EQ number 33u64)"), 2);
+        // Doc 30 of the first segment is deleted.
+        assert_eq!(count("(EQ number 30u64)"), 1);
+        // `label` is "a" on 25 docs of the second segment, one of which (1060) is deleted.
+        assert_eq!(count(r#"(EQ label "a")"#), 24);
+    }
+
+    #[test]
+    fn test_required_presence_restricts_evaluated_docs() {
+        let index = create_sparse_index();
+        let searcher = index.reader().unwrap().searcher();
+        let segment_reader = searcher.segment_reader(0);
+        let size_hint = |query: DocPredicateQuery| {
+            query
+                .weight(EnableScoring::disabled_from_searcher(&searcher))
+                .unwrap()
+                .scorer(segment_reader, 1.0)
+                .unwrap()
+                .size_hint()
+        };
+        // `number` has a value in one doc out of three.
+        assert_eq!(size_hint(query("(GT number 10u64)")), 100);
+        assert_eq!(
+            size_hint(query_without_required_presence("(GT number 10u64)")),
+            300
+        );
+        // Nothing to require: all docs are evaluated.
+        assert_eq!(size_hint(query("(NOT (GT number 10u64))")), 300);
+    }
+
+    // THIS FAILS! due to our pick best possible column approach policy.
+    // #[test]
+    // fn test_multi_typed_field_picks_one() {
+    //     let mut schema_builder = Schema::builder();
+    //     let json = schema_builder.add_json_field("json", FAST);
+    //     let index = Index::create_in_ram(schema_builder.build());
+    //     let mut writer = index.writer_for_tests().unwrap();
+    //     writer
+    //         .add_document(doc!(json => serde_json::json!({"myfield": 2u64})))
+    //         .unwrap();
+    //     writer
+    //         .add_document(doc!(json => serde_json::json!({"myfield": "b"})))
+    //         .unwrap();
+    //     writer.commit().unwrap();
+    //     let searcher = index.reader().unwrap().searcher();
+    //     assert_eq!(
+    //         searcher
+    //             .search(&query(r#"(IS_NULL json.myfield)"#), &Count)
+    //             .unwrap(),
+    //         2 // assertion fails, expected 2 got 1
+    //     );
+    // }
+
+    // THIS FAILS DUE TO EQ infer_types being too lenient.
+    // #[test]
+    // fn test_multi_typed_field_eq_too_lenient_failing() {
+    //     let mut schema_builder = Schema::builder();
+    //     let json = schema_builder.add_json_field("json", FAST);
+    //     let index = Index::create_in_ram(schema_builder.build());
+    //     let mut writer = index.writer_for_tests().unwrap();
+    //     writer
+    //         .add_document(doc!(json => serde_json::json!({"myfield": 2u64})))
+    //         .unwrap();
+    //     writer
+    //         .add_document(doc!(json => serde_json::json!({"myfield": "b"})))
+    //         .unwrap();
+    //     writer.commit().unwrap();
+    //     let searcher = index.reader().unwrap().searcher();
+    //     assert_eq!(
+    //         searcher
+    //             .search(&query(r#"(EQ json.myfield "b")"#), &Count)
+    //             .unwrap(),
+    //         1 // assertion fails, expected 2 got 1
+    //     );
+    // }
+}
diff --git a/src/query/doc_predicate_query/mod.rs b/src/query/doc_predicate_query/mod.rs
index 5346253de..e8c74e106 100644
--- a/src/query/doc_predicate_query/mod.rs
+++ b/src/query/doc_predicate_query/mod.rs
@@ -1,13 +1,20 @@
+use std::cmp::Ordering;
 use std::sync::Arc;
 
 mod function_predicate;
+#[cfg(feature = "jitexpr")]
+mod jitexpr_predicate;
 
 pub use function_predicate::FunctionPredicate;
+#[cfg(feature = "jitexpr")]
+pub use jitexpr_predicate::{JitExprEvalState, JitExprPredicate};
 
 use crate::docset::{SeekDangerResult, TERMINATED};
 use crate::index::SegmentReader;
 use crate::query::explanation::does_not_match;
-use crate::query::{ConstScorer, EnableScoring, Explanation, Query, Scorer, Weight};
+use crate::query::{
+    AllWeight, ConstScorer, EmptyWeight, EnableScoring, Explanation, Query, Scorer, Weight,
+};
 use crate::{DocId, DocSet, Score};
 
 /// A query that evaluates, for each DocId, whether it matches or not.
@@ -59,67 +66,130 @@ impl Weight for DocPredicateQuery {
     }
 }
 
-/// A [`DocSet`] that walks documents by repeatedly evaluating a
-/// [`SegmentDocPredicate`], starting from doc `0`.
+/// A [`DocSet`] that walks the documents of a necessary condition, and evaluates a
+/// [`SegmentDocPredicate`] on each of them.
+///
+/// Hidden contract: every document matching the predicate belongs to the necessary condition.
+/// Documents outside of it are never evaluated, and are considered as not matching.
+///
+/// Hidden contract: whenever the `DocPredicateDocSet` is in a valid state, the necessary condition
+/// is in a valid state too, positioned on a matching document (or `TERMINATED`). The current doc
+/// is therefore simply the necessary condition's current doc.
 pub struct DocPredicateDocSet<TSegmentDocPredicate> {
     doc_predicate: TSegmentDocPredicate,
-    doc: DocId,
-    max_doc: DocId,
+    necessary_condition: Box<dyn DocSet>,
+}
+
+impl<TSegmentDocPredicate: SegmentDocPredicate> DocPredicateDocSet<TSegmentDocPredicate> {
+    /// Creates a `DocPredicateDocSet` positioned on its first matching document.
+    fn new(doc_predicate: TSegmentDocPredicate, necessary_condition: Box<dyn DocSet>) -> Self {
+        let first_candidate = necessary_condition.doc();
+        let mut doc_set = DocPredicateDocSet {
+            doc_predicate,
+            necessary_condition,
+        };
+        doc_set.find_match(first_candidate);
+        doc_set
+    }
+
+    /// Creates a `DocPredicateDocSet`, and seeks it to `target`, following
+    /// [`Weight::scorer_danger`]'s contract.
+    ///
+    /// Documents before `target` are not evaluated.
+    fn new_seeked_to(
+        doc_predicate: TSegmentDocPredicate,
+        necessary_condition: Box<dyn DocSet>,
+        target: DocId,
+    ) -> (SeekDangerResult, Self) {
+        let first_candidate = necessary_condition.doc();
+        let mut doc_set = DocPredicateDocSet {
+            doc_predicate,
+            necessary_condition,
+        };
+        if target >= TERMINATED {
+            if doc_set.necessary_condition.doc() < TERMINATED {
+                doc_set.necessary_condition.seek(TERMINATED);
+            }
+            return (SeekDangerResult::SeekLowerBound(TERMINATED), doc_set);
+        }
+        let seek_result = match first_candidate.cmp(&target) {
+            Ordering::Less => doc_set.seek_danger(target),
+            Ordering::Equal => doc_set.eval_candidate(target),
+            Ordering::Greater => SeekDangerResult::SeekLowerBound(first_candidate),
+        };
+        (seek_result, doc_set)
+    }
+
+    /// Evaluates the predicate on `candidate`.
+    ///
+    /// Hidden contract: the necessary condition is positioned on `candidate`.
+    fn eval_candidate(&mut self, candidate: DocId) -> SeekDangerResult {
+        if self.doc_predicate.eval(candidate) {
+            SeekDangerResult::Found
+        } else {
+            SeekDangerResult::SeekLowerBound(candidate + 1)
+        }
+    }
+
+    /// Advances to the first matching document at or after `candidate`.
+    ///
+    /// Hidden contract: the necessary condition is positioned on `candidate`.
+    fn find_match(&mut self, mut candidate: DocId) -> DocId {
+        debug_assert_eq!(candidate, self.necessary_condition.doc());
+        while candidate != TERMINATED && !self.doc_predicate.eval(candidate) {
+            candidate = self.necessary_condition.advance();
+        }
+        candidate
+    }
 }
 
 impl<TSegmentDocPredicate: SegmentDocPredicate> DocSet
     for DocPredicateDocSet<TSegmentDocPredicate>
 {
     fn advance(&mut self) -> DocId {
-        if self.doc == TERMINATED {
+        if self.doc() == TERMINATED {
             return TERMINATED;
         }
-        self.find_match(self.doc + 1)
+        let candidate = self.necessary_condition.advance();
+        self.find_match(candidate)
     }
 
     fn seek(&mut self, target: DocId) -> DocId {
-        debug_assert!(target >= self.doc);
-        if self.doc == TERMINATED {
-            return TERMINATED;
+        let doc = self.doc();
+        debug_assert!(target >= doc);
+        // In a valid state, the current doc is a match (or TERMINATED).
+        if doc >= target {
+            return doc;
         }
-        self.find_match(target)
+        let candidate = self.necessary_condition.seek(target);
+        self.find_match(candidate)
     }
 
     fn seek_danger(&mut self, target: DocId) -> SeekDangerResult {
-        if target >= self.max_doc {
-            self.doc = TERMINATED;
-            return SeekDangerResult::SeekLowerBound(TERMINATED);
-        }
-        if self.doc_predicate.eval(target) {
-            self.doc = target;
-            SeekDangerResult::Found
-        } else {
-            SeekDangerResult::SeekLowerBound(target + 1)
+        match self.necessary_condition.seek_danger(target) {
+            SeekDangerResult::Found => self.eval_candidate(target),
+            // Following `seek_danger`'s contract, we are now in an invalid state, and `doc()` may
+            // return anything until a subsequent `seek_danger` returns `Found`.
+            seek_lower_bound @ SeekDangerResult::SeekLowerBound(_) => seek_lower_bound,
         }
     }
 
     fn doc(&self) -> DocId {
-        self.doc
+        self.necessary_condition.doc()
     }
 
     fn size_hint(&self) -> u32 {
-        self.max_doc
+        self.necessary_condition.size_hint()
     }
-}
 
-impl<TSegmentDocPredicate: SegmentDocPredicate> DocPredicateDocSet<TSegmentDocPredicate> {
-    fn find_match(&mut self, mut target: DocId) -> DocId {
-        loop {
-            match self.seek_danger(target) {
-                SeekDangerResult::Found => return target,
-                SeekDangerResult::SeekLowerBound(next_target) => {
-                    if next_target >= TERMINATED {
-                        return TERMINATED;
-                    }
-                    target = next_target;
-                }
-            }
-        }
+    fn cost(&self) -> u64 {
+        // `cost` is the method used to tell  how costly it is to consume a DocSet entirely.
+        //
+        // This is used in intersection to have cheaper docset "lead" the intersection.
+        //
+        // Here, we naturally use a model where we use the cost of the necessary condition
+        // multiplied by some factor expressing how slow it is to evaluate an expression.
+        self.necessary_condition.cost() * self.doc_predicate.cost()
     }
 }
 
@@ -141,14 +211,27 @@ pub trait DocPredicateBoxable: std::fmt::Debug + 'static + Send + Sync {
 
 impl<TDocPredicate: DocPredicate> DocPredicateBoxable for TDocPredicate {
     fn scorer(&self, segment_reader: &SegmentReader, boost: f32) -> crate::Result<Box<dyn Scorer>> {
-        let doc_predicate = self.doc_predicate(segment_reader)?;
-        let mut doc_set = DocPredicateDocSet {
-            doc_predicate,
-            doc: 0u32,
-            max_doc: segment_reader.max_doc(),
-        };
-        doc_set.doc = doc_set.find_match(0);
-        Ok(Box::new(ConstScorer::new(doc_set, boost)) as Box<dyn Scorer>)
+        let const_or_variable_segment_predicate = self.doc_predicate(segment_reader)?;
+        match const_or_variable_segment_predicate {
+            ConstOrVariableSegmentPredicate::Const(always_match) => {
+                if always_match {
+                    AllWeight.scorer(segment_reader, boost)
+                } else {
+                    EmptyWeight.scorer(segment_reader, boost)
+                }
+            }
+            ConstOrVariableSegmentPredicate::Variable {
+                predicate,
+                necessary_condition,
+            } => {
+                if necessary_condition.doc() >= segment_reader.max_doc() {
+                    // The necessary condition is empty.
+                    return EmptyWeight.scorer(segment_reader, boost);
+                }
+                let doc_set = DocPredicateDocSet::new(predicate, necessary_condition);
+                Ok(Box::new(ConstScorer::new(doc_set, boost)))
+            }
+        }
     }
 
     fn scorer_danger(
@@ -157,18 +240,48 @@ impl<TDocPredicate: DocPredicate> DocPredicateBoxable for TDocPredicate {
         target: DocId,
         boost: f32,
     ) -> crate::Result<(SeekDangerResult, Box<dyn Scorer>)> {
-        let doc_predicate = self.doc_predicate(segment_reader)?;
-        let mut doc_set = DocPredicateDocSet {
-            doc_predicate,
-            doc: target,
-            max_doc: segment_reader.max_doc(),
-        };
-        let seek_result = doc_set.seek_danger(target);
-        let scorer = Box::new(ConstScorer::new(doc_set, boost)) as Box<dyn Scorer>;
-        Ok((seek_result, scorer))
+        let const_or_variable_segment_predicate = self.doc_predicate(segment_reader)?;
+        match const_or_variable_segment_predicate {
+            ConstOrVariableSegmentPredicate::Const(always_match) => {
+                if always_match {
+                    AllWeight.scorer_danger(segment_reader, target, boost)
+                } else {
+                    EmptyWeight.scorer_danger(segment_reader, target, boost)
+                }
+            }
+            ConstOrVariableSegmentPredicate::Variable {
+                predicate,
+                necessary_condition,
+            } => {
+                let (seek_result, doc_set) =
+                    DocPredicateDocSet::new_seeked_to(predicate, necessary_condition, target);
+                Ok((seek_result, Box::new(ConstScorer::new(doc_set, boost))))
+            }
+        }
     }
 }
 
+/// Represents a segment predicate.
+pub enum ConstOrVariableSegmentPredicate<P: SegmentDocPredicate> {
+    /// Can be emitted to hint that a predicate will be always true or false on a segment.
+    /// Returning Const instead of a variable is an optimization.
+    Const(bool),
+    /// A regular SegmentDocPredicate, evaluated document by document.
+    Variable {
+        /// The predicate to evaluate.
+        predicate: P,
+        /// The [`DocSet`] of the documents on which `predicate` is evaluated.
+        ///
+        /// Hidden contract: it must contain every document for which `predicate.eval` returns
+        /// true. Documents outside of it are never evaluated, and are considered as not
+        /// matching. Use an [`AllScorer`](crate::query::AllScorer) to evaluate every document of
+        /// the segment.
+        ///
+        /// The `DocSet` must be positioned on its first document.
+        necessary_condition: Box<dyn DocSet>,
+    },
+}
+
 /// A per-query predicate that produces a [`SegmentDocPredicate`] for each
 /// segment.
 ///
@@ -186,19 +299,37 @@ pub trait DocPredicate: Send + Sync + 'static + std::fmt::Debug {
     fn doc_predicate(
         &self,
         segment_reader: &SegmentReader,
-    ) -> crate::Result<Self::SegmentDocPredicate>;
+    ) -> crate::Result<ConstOrVariableSegmentPredicate<Self::SegmentDocPredicate>>;
 }
 
 /// The per-segment predicate produced by a [`DocPredicate`].
 pub trait SegmentDocPredicate: Send + 'static {
     /// Returns whether `doc_id` matches the predicate.
     fn eval(&mut self, doc_id: DocId) -> bool;
+
+    /// Cost for the evaluation of a given predicate.
+    ///
+    /// This is used to infer the cost of consuming an associated `DocPredicateDocSet`.
+    /// This does not need to be accurate. It is only used by the intersection scorer
+    /// to choose which `DocSet` should "drive" the intersection.
+    ///
+    /// 1 is the time it takes to call `TermScorer::advance` (a few cycles). We defensively default
+    /// to 100.
+    fn cost(&self) -> u64 {
+        // We assume a default value of 100.
+        100u64
+    }
 }
 
 #[cfg(test)]
 pub(crate) mod tests {
+    use std::sync::atomic::{AtomicUsize, Ordering as AtomicOrdering};
+
+    use proptest::prelude::*;
+
     use super::*;
-    use crate::collector::Count;
+    use crate::collector::{Count, DocSetCollector};
+    use crate::query::VecDocSet;
 
     pub(crate) fn create_index_for_test(num_docs: u32) -> crate::Index {
         let schema_builder = crate::schema::Schema::builder();
@@ -271,6 +402,207 @@ pub(crate) mod tests {
         assert_eq!(scorer.doc(), 2);
     }
 
+    /// Matches even doc ids, and counts its evaluations.
+    struct EvenDocIds {
+        num_evals: Arc<AtomicUsize>,
+    }
+
+    impl SegmentDocPredicate for EvenDocIds {
+        fn eval(&mut self, doc_id: DocId) -> bool {
+            self.num_evals.fetch_add(1, AtomicOrdering::Relaxed);
+            doc_id.is_multiple_of(2)
+        }
+    }
+
+    /// `EvenDocIds`, with a fixed necessary condition.
+    #[derive(Debug)]
+    struct EvenWithNecessaryCondition {
+        necessary_condition: Vec<DocId>,
+        num_evals: Arc<AtomicUsize>,
+    }
+
+    impl EvenWithNecessaryCondition {
+        fn new(necessary_condition: Vec<DocId>) -> Self {
+            EvenWithNecessaryCondition {
+                necessary_condition,
+                num_evals: Arc::default(),
+            }
+        }
+    }
+
+    impl DocPredicate for EvenWithNecessaryCondition {
+        type SegmentDocPredicate = EvenDocIds;
+
+        fn doc_predicate(
+            &self,
+            _segment_reader: &SegmentReader,
+        ) -> crate::Result<ConstOrVariableSegmentPredicate<EvenDocIds>> {
+            Ok(ConstOrVariableSegmentPredicate::Variable {
+                predicate: EvenDocIds {
+                    num_evals: self.num_evals.clone(),
+                },
+                necessary_condition: Box::new(VecDocSet::from(self.necessary_condition.clone())),
+            })
+        }
+    }
+
+    #[test]
+    fn test_necessary_condition_restricts_evaluations() {
+        let index = create_index_for_test(10);
+        let searcher = index.reader().unwrap().searcher();
+        let predicate = EvenWithNecessaryCondition::new(vec![1, 2, 3, 4, 6, 9]);
+        let num_evals = predicate.num_evals.clone();
+        let query: DocPredicateQuery = predicate.into();
+        assert_eq!(searcher.search(&query, &DocSetCollector).unwrap().len(), 3);
+        assert_eq!(num_evals.load(AtomicOrdering::Relaxed), 6);
+    }
+
+    #[test]
+    fn test_necessary_condition_size_hint_and_cost() {
+        let index = create_index_for_test(10);
+        let searcher = index.reader().unwrap().searcher();
+        let query: DocPredicateQuery =
+            EvenWithNecessaryCondition::new(vec![1, 2, 3, 4, 6, 9]).into();
+        let weight = query
+            .weight(EnableScoring::disabled_from_searcher(&searcher))
+            .unwrap();
+        let scorer = weight.scorer(searcher.segment_reader(0), 1.0).unwrap();
+        assert_eq!(scorer.size_hint(), 6);
+        assert_eq!(scorer.cost(), 600);
+        // Without a necessary condition, all docs are candidates.
+        let scorer = even_doc_id_query()
+            .scorer(searcher.segment_reader(0), 1.0)
+            .unwrap();
+        assert_eq!(scorer.size_hint(), 10);
+        assert_eq!(scorer.cost(), 1000);
+    }
+
+    #[test]
+    fn test_necessary_condition_scorer_danger() {
+        let index = create_index_for_test(10);
+        let searcher = index.reader().unwrap().searcher();
+        let segment_reader = searcher.segment_reader(0);
+        let scorer_danger = |necessary_condition: Vec<DocId>, target: DocId| {
+            let predicate = EvenWithNecessaryCondition::new(necessary_condition);
+            let num_evals = predicate.num_evals.clone();
+            let query: DocPredicateQuery = predicate.into();
+            let (seek_result, scorer) = query.scorer_danger(segment_reader, target, 1.0).unwrap();
+            (seek_result, scorer, num_evals.load(AtomicOrdering::Relaxed))
+        };
+
+        // The necessary condition starts after the target.
+        let (seek_result, mut scorer, num_evals) = scorer_danger(vec![4, 6], 1);
+        assert_eq!(seek_result, SeekDangerResult::SeekLowerBound(4));
+        assert_eq!(num_evals, 0);
+        assert_eq!(scorer.seek_danger(4), SeekDangerResult::Found);
+        assert_eq!(scorer.doc(), 4);
+
+        // The target is the first candidate, and matches.
+        let (seek_result, scorer, _) = scorer_danger(vec![2, 6], 2);
+        assert_eq!(seek_result, SeekDangerResult::Found);
+        assert_eq!(scorer.doc(), 2);
+
+        // The target is a candidate, but does not match.
+        let (seek_result, mut scorer, num_evals) = scorer_danger(vec![1, 3, 4], 3);
+        assert_eq!(seek_result, SeekDangerResult::SeekLowerBound(4));
+        assert_eq!(num_evals, 1);
+        assert_eq!(scorer.seek_danger(4), SeekDangerResult::Found);
+
+        // The target is not a candidate: it is not evaluated.
+        let (seek_result, _, num_evals) = scorer_danger(vec![1, 6], 2);
+        assert_eq!(seek_result, SeekDangerResult::SeekLowerBound(6));
+        assert_eq!(num_evals, 0);
+
+        // No match after the target. The lower bound can stop on a non-matching candidate.
+        let (seek_result, mut scorer, _) = scorer_danger(vec![1, 3], 2);
+        assert_eq!(seek_result, SeekDangerResult::SeekLowerBound(3));
+        assert_eq!(scorer.seek_danger(3), SeekDangerResult::SeekLowerBound(4));
+        assert_eq!(
+            scorer.seek_danger(4),
+            SeekDangerResult::SeekLowerBound(TERMINATED)
+        );
+    }
+
+    proptest! {
+        #[test]
+        fn proptest_necessary_condition_doc_set(
+            candidates in prop::collection::btree_set(0u32..200, 0..60),
+            modulo in 1u32..5,
+            targets in prop::collection::vec(0u32..220, 0..30),
+            advances in prop::collection::vec(any::<bool>(), 0..30),
+        ) {
+            let candidates: Vec<DocId> = candidates.into_iter().collect();
+            let expected: Vec<DocId> = candidates
+                .iter()
+                .copied()
+                .filter(|doc| doc.is_multiple_of(modulo))
+                .collect();
+            let new_doc_set = || {
+                DocPredicateDocSet::new(
+                    move |doc: DocId| doc.is_multiple_of(modulo),
+                    Box::new(VecDocSet::from(candidates.clone())),
+                )
+            };
+            let first_match = |target: DocId| {
+                expected
+                    .iter()
+                    .copied()
+                    .find(|doc| *doc >= target)
+                    .unwrap_or(TERMINATED)
+            };
+
+            // advance
+            let mut doc_set = new_doc_set();
+            let mut matches: Vec<DocId> = Vec::new();
+            while doc_set.doc() != TERMINATED {
+                matches.push(doc_set.doc());
+                doc_set.advance();
+            }
+            prop_assert_eq!(&matches, &expected);
+
+            // interleaved seek and advance
+            let mut doc_set = new_doc_set();
+            for (target, advance) in targets.iter().zip(advances.iter()) {
+                let target = (*target).max(doc_set.doc());
+                if *advance && doc_set.doc() != TERMINATED {
+                    let current = doc_set.doc();
+                    prop_assert_eq!(doc_set.advance(), first_match(current + 1));
+                } else {
+                    prop_assert_eq!(doc_set.seek(target), first_match(target));
+                }
+            }
+
+            // seek_danger, following its contract: strictly increasing targets, respecting
+            // the returned lower bounds.
+            let mut sorted_targets = targets.clone();
+            sorted_targets.sort_unstable();
+            sorted_targets.dedup();
+            let mut doc_set = new_doc_set();
+            let mut lower_bound = doc_set.doc();
+            let mut previous_target = None;
+            for requested_target in sorted_targets {
+                let target = requested_target.max(lower_bound);
+                if previous_target.is_some_and(|previous| previous >= target) {
+                    continue;
+                }
+                previous_target = Some(target);
+                let next_match = first_match(target);
+                match doc_set.seek_danger(target) {
+                    SeekDangerResult::Found => {
+                        prop_assert_eq!(next_match, target);
+                        prop_assert_eq!(doc_set.doc(), target);
+                    }
+                    SeekDangerResult::SeekLowerBound(bound) => {
+                        prop_assert!(next_match != target || target == TERMINATED);
+                        prop_assert!(bound > target || target == TERMINATED);
+                        prop_assert!(bound <= next_match);
+                        lower_bound = bound;
+                    }
+                }
+            }
+        }
+    }
+
     #[test]
     fn test_doc_predicate_query_scorer_danger_target_past_max_doc() {
         let index = create_index_for_test(4);
diff --git a/src/query/exist_query.rs b/src/query/exist_query.rs
index fcda85fff..a0121fbe9 100644
--- a/src/query/exist_query.rs
+++ b/src/query/exist_query.rs
@@ -182,7 +182,7 @@ impl Weight for FastFieldExistsWeight {
     }
 }
 
-enum ExistsColumnIndex {
+pub(crate) enum ExistsColumnIndex {
     Optional(OptionalIndex),
     Multivalued(MultiValueIndex),
 }
diff --git a/src/query/fuzzy_query.rs b/src/query/fuzzy_query.rs
index a0634b96b..a3f8faa43 100644
--- a/src/query/fuzzy_query.rs
+++ b/src/query/fuzzy_query.rs
@@ -6,7 +6,11 @@ use crate::query::{AutomatonWeight, EnableScoring, Query, Weight};
 use crate::schema::{Term, Type};
 use crate::TantivyError::InvalidArgument;
 
-pub(crate) struct DfaWrapper(pub DFA);
+/// The Levenshtein automaton a [`FuzzyTermQuery`] matches terms with.
+///
+/// Obtained from [`FuzzyTermQuery::automaton`], e.g. to stream the term dictionary
+/// and collect the terms the query expands to.
+pub struct DfaWrapper(pub(crate) DFA);
 
 impl Automaton for DfaWrapper {
     type State = u32;
@@ -109,7 +113,21 @@ impl FuzzyTermQuery {
         }
     }
 
-    fn specialized_weight(&self) -> crate::Result<AutomatonWeight<DfaWrapper>> {
+    /// Returns the automaton this query matches terms with.
+    ///
+    /// For a JSON term, it matches the term's text only, not its JSON path.
+    ///
+    /// ```rust
+    /// use tantivy::query::{DfaWrapper, FuzzyTermQuery};
+    /// use tantivy::schema::{Schema, TEXT};
+    /// use tantivy::Term;
+    ///
+    /// let mut schema_builder = Schema::builder();
+    /// let title = schema_builder.add_text_field("title", TEXT);
+    /// let query = FuzzyTermQuery::new(Term::from_field_text(title, "diary"), 1, true);
+    /// let _automaton: DfaWrapper = query.automaton().unwrap();
+    /// ```
+    pub fn automaton(&self) -> crate::Result<DfaWrapper> {
         static AUTOMATON_BUILDER: [[OnceCell<LevenshteinAutomatonBuilder>; 2]; 3] = [
             [OnceCell::new(), OnceCell::new()],
             [OnceCell::new(), OnceCell::new()],
@@ -158,18 +176,19 @@ impl FuzzyTermQuery {
         } else {
             automaton_builder.build_dfa(term_text)
         };
+        Ok(DfaWrapper(automaton))
+    }
 
-        if let Some((json_path_bytes, _)) = term_value.as_json() {
+    fn specialized_weight(&self) -> crate::Result<AutomatonWeight<DfaWrapper>> {
+        let automaton = self.automaton()?;
+        if let Some((json_path_bytes, _)) = self.term.value().as_json() {
             Ok(AutomatonWeight::new_for_json_path(
                 self.term.field(),
-                DfaWrapper(automaton),
+                automaton,
                 json_path_bytes,
             ))
         } else {
-            Ok(AutomatonWeight::new(
-                self.term.field(),
-                DfaWrapper(automaton),
-            ))
+            Ok(AutomatonWeight::new(self.term.field(), automaton))
         }
     }
 }
@@ -189,6 +208,30 @@ mod test {
     use crate::schema::{Schema, STORED, TEXT};
     use crate::{assert_nearly_equals, Index, IndexWriter, TantivyDocument, Term};
 
+    #[test]
+    pub fn test_fuzzy_automaton_streams_expanded_terms() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let title = schema_builder.add_text_field("title", TEXT);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut index_writer: IndexWriter = index.writer_for_tests()?;
+        index_writer.add_document(doc!(title => "The Diary of a Dairy Cow"))?;
+        index_writer.add_document(doc!(title => "A Daily Log"))?;
+        index_writer.commit()?;
+        let searcher = index.reader()?.searcher();
+
+        let query = FuzzyTermQuery::new(Term::from_field_text(title, "diary"), 1, true);
+        let automaton = query.automaton()?;
+        let inverted_index = searcher.segment_reader(0).inverted_index(title)?;
+        let mut stream = inverted_index.terms().search(automaton).into_stream()?;
+        let mut terms = Vec::new();
+        while stream.advance() {
+            terms.push(String::from_utf8(stream.key().to_vec()).unwrap());
+        }
+        assert_eq!(terms, vec!["dairy", "diary"]);
+        assert_eq!(searcher.search(&query, &Count)?, 1);
+        Ok(())
+    }
+
     #[test]
     pub fn test_fuzzy_json_path() -> crate::Result<()> {
         // # Defining the schema
diff --git a/src/query/mod.rs b/src/query/mod.rs
index a1189c793..dd168e03b 100644
--- a/src/query/mod.rs
+++ b/src/query/mod.rs
@@ -51,9 +51,7 @@ pub use self::empty_query::{EmptyQuery, EmptyScorer, EmptyWeight};
 pub use self::exclude::{Exclude, ExclusionSet};
 pub use self::exist_query::ExistsQuery;
 pub use self::explanation::Explanation;
-#[cfg(test)]
-pub(crate) use self::fuzzy_query::DfaWrapper;
-pub use self::fuzzy_query::FuzzyTermQuery;
+pub use self::fuzzy_query::{DfaWrapper, FuzzyTermQuery};
 pub use self::intersection::{intersect_scorers, Intersection};
 pub use self::more_like_this::{MoreLikeThisQuery, MoreLikeThisQueryBuilder};
 pub use self::phrase_prefix_query::PhrasePrefixQuery;
diff --git a/src/query/phrase_query/regex_phrase_query.rs b/src/query/phrase_query/regex_phrase_query.rs
index 98e07d399..67ca03499 100644
--- a/src/query/phrase_query/regex_phrase_query.rs
+++ b/src/query/phrase_query/regex_phrase_query.rs
@@ -1,3 +1,9 @@
+use std::fmt;
+use std::sync::Arc;
+
+use once_cell::sync::OnceCell;
+use tantivy_fst::Regex;
+
 use super::regex_phrase_weight::RegexPhraseWeight;
 use crate::query::bm25::Bm25Weight;
 use crate::query::{EnableScoring, Query, Weight};
@@ -25,6 +31,21 @@ pub struct RegexPhraseQuery {
     phrase_terms: Vec<(usize, String)>,
     slop: u32,
     max_expansions: u32,
+    regexes: CompiledRegexes,
+}
+
+/// The compiled `phrase_terms`, built on first use. Its `Debug` omits the automata.
+#[derive(Clone, Default)]
+struct CompiledRegexes(OnceCell<Vec<Arc<Regex>>>);
+
+impl fmt::Debug for CompiledRegexes {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.write_str(if self.0.get().is_some() {
+            "CompiledRegexes(compiled)"
+        } else {
+            "CompiledRegexes(pending)"
+        })
+    }
 }
 
 /// Transform a wildcard query to a regex string.
@@ -71,6 +92,37 @@ impl RegexPhraseQuery {
             phrase_terms: terms,
             slop,
             max_expansions: 1 << 14,
+            regexes: CompiledRegexes::default(),
+        }
+    }
+
+    /// Creates a new `RegexPhraseQuery` from already compiled regexes, e.g. built with a
+    /// non-default state limit.
+    ///
+    /// Each term is `(offset, pattern, regex)`, where `regex` is the compilation of
+    /// `pattern`; the pattern is what [`RegexPhraseQuery::phrase_terms`] returns.
+    pub fn from_regexes(
+        field: Field,
+        mut terms: Vec<(usize, String, Arc<Regex>)>,
+        slop: u32,
+    ) -> RegexPhraseQuery {
+        assert!(
+            terms.len() > 1,
+            "A phrase query is required to have strictly more than one term."
+        );
+        terms.sort_by_key(|&(offset, _, _)| offset);
+        let (phrase_terms, regexes): (Vec<_>, Vec<_>) = terms
+            .into_iter()
+            .map(|(offset, pattern, regex)| ((offset, pattern), regex))
+            .unzip();
+        let compiled = OnceCell::new();
+        let _ = compiled.set(regexes);
+        RegexPhraseQuery {
+            field,
+            phrase_terms,
+            slop,
+            max_expansions: 1 << 14,
+            regexes: CompiledRegexes(compiled),
         }
     }
 
@@ -111,6 +163,24 @@ impl RegexPhraseQuery {
             .collect::<Vec<Term>>()
     }
 
+    /// The compiled regex of each phrase term, in offset order.
+    ///
+    /// Compiled on first call and cached, so a query that is reused across searchers,
+    /// or inspected before searching, determinizes each pattern once.
+    pub fn regexes(&self) -> crate::Result<&[Arc<Regex>]> {
+        let regexes = self.regexes.0.get_or_try_init(|| {
+            self.phrase_terms
+                .iter()
+                .map(|(_, term)| {
+                    Regex::new(term).map(Arc::new).map_err(|e| {
+                        crate::TantivyError::InvalidArgument(format!("Invalid regex: {e}"))
+                    })
+                })
+                .collect::<crate::Result<Vec<_>>>()
+        })?;
+        Ok(regexes)
+    }
+
     /// Returns the [`RegexPhraseWeight`] for the given phrase query given a specific `searcher`.
     ///
     /// This function is the same as [`Query::weight()`] except it returns
@@ -149,9 +219,15 @@ impl RegexPhraseQuery {
             } => Some(Bm25Weight::for_terms(statistics_provider, &terms)?),
             EnableScoring::Disabled { .. } => None,
         };
+        let phrase_terms = self
+            .phrase_terms
+            .iter()
+            .map(|(offset, _)| *offset)
+            .zip(self.regexes()?.iter().cloned())
+            .collect();
         let weight = RegexPhraseWeight::new(
             self.field,
-            self.phrase_terms.clone(),
+            phrase_terms,
             bm25_weight_opt,
             self.max_expansions,
             self.slop,
diff --git a/src/query/phrase_query/regex_phrase_weight.rs b/src/query/phrase_query/regex_phrase_weight.rs
index 9cefc555a..bbaddbc79 100644
--- a/src/query/phrase_query/regex_phrase_weight.rs
+++ b/src/query/phrase_query/regex_phrase_weight.rs
@@ -20,7 +20,7 @@ type UnionType = SimpleUnion<Box<dyn Postings + 'static>>;
 /// See RegexPhraseWeight::get_union_from_term_infos for some design decisions.
 pub struct RegexPhraseWeight {
     field: Field,
-    phrase_terms: Vec<(usize, String)>,
+    phrase_terms: Vec<(usize, Arc<Regex>)>,
     similarity_weight_opt: Option<Bm25Weight>,
     slop: u32,
     max_expansions: u32,
@@ -31,7 +31,7 @@ impl RegexPhraseWeight {
     /// If `similarity_weight_opt` is None, then scoring is disabled
     pub fn new(
         field: Field,
-        phrase_terms: Vec<(usize, String)>,
+        phrase_terms: Vec<(usize, Arc<Regex>)>,
         similarity_weight_opt: Option<Bm25Weight>,
         max_expansions: u32,
         slop: u32,
@@ -67,12 +67,9 @@ impl RegexPhraseWeight {
         let mut posting_lists = Vec::new();
         let inverted_index = reader.inverted_index(self.field)?;
         let mut num_terms = 0;
-        for &(offset, ref term) in &self.phrase_terms {
-            let regex = Regex::new(term)
-                .map_err(|e| crate::TantivyError::InvalidArgument(format!("Invalid regex: {e}")))?;
-
+        for &(offset, ref regex) in &self.phrase_terms {
             let automaton: AutomatonWeight<Regex> =
-                AutomatonWeight::new(self.field, Arc::new(regex));
+                AutomatonWeight::new(self.field, Arc::clone(regex));
             let term_infos = automaton.get_match_term_infos(reader)?;
             // If term_infos is empty, the phrase can not match any documents.
             if term_infos.is_empty() {
@@ -351,6 +348,71 @@ mod tests {
         }
     }
 
+    #[test]
+    pub fn test_phrase_regex_invalid_pattern_fails_at_weight() -> crate::Result<()> {
+        let index = create_index(&["a b"])?;
+        let text_field = index.schema().get_field("text").unwrap();
+        let searcher = index.reader()?.searcher();
+        let phrase_query = RegexPhraseQuery::new(text_field, vec!["a".into(), "(".into()]);
+        let enable_scoring = EnableScoring::enabled_from_searcher(&searcher);
+        assert!(matches!(
+            phrase_query.regex_phrase_weight(enable_scoring),
+            Err(crate::TantivyError::InvalidArgument(_))
+        ));
+        Ok(())
+    }
+
+    #[test]
+    pub fn test_phrase_regexes_are_compiled_once_and_shared() -> crate::Result<()> {
+        let index = create_index(&["a b"])?;
+        let text_field = index.schema().get_field("text").unwrap();
+        let searcher = index.reader()?.searcher();
+        let phrase_query = RegexPhraseQuery::new(text_field, vec!["a.*".into(), "b".into()]);
+        let first = phrase_query.regexes()?.as_ptr();
+        assert_eq!(phrase_query.regexes()?.as_ptr(), first);
+
+        let enable_scoring = EnableScoring::enabled_from_searcher(&searcher);
+        let _weight = phrase_query.regex_phrase_weight(enable_scoring)?;
+        let clone = phrase_query.clone();
+        let _clone_weight = clone.regex_phrase_weight(enable_scoring)?;
+        // The query, its clone and both weights hold the same automata.
+        for regex in phrase_query.regexes()? {
+            assert_eq!(std::sync::Arc::strong_count(regex), 4);
+        }
+        Ok(())
+    }
+
+    #[test]
+    pub fn test_phrase_from_regexes_uses_the_given_automata() -> crate::Result<()> {
+        use std::sync::Arc;
+
+        use tantivy_fst::Regex;
+
+        use crate::collector::Count;
+
+        let index = create_index(&["a b", "aa b", "b a", "a c"])?;
+        let text_field = index.schema().get_field("text").unwrap();
+        let searcher = index.reader()?.searcher();
+        let regex_a = Arc::new(Regex::new("a.*").unwrap());
+        let regex_b = Arc::new(Regex::new("b").unwrap());
+        let from_regexes = RegexPhraseQuery::from_regexes(
+            text_field,
+            vec![
+                (1, "b".into(), regex_b.clone()),
+                (0, "a.*".into(), regex_a.clone()),
+            ],
+            0,
+        );
+        let regexes = from_regexes.regexes()?;
+        assert!(Arc::ptr_eq(&regexes[0], &regex_a));
+        assert!(Arc::ptr_eq(&regexes[1], &regex_b));
+
+        let from_patterns = RegexPhraseQuery::new(text_field, vec!["a.*".into(), "b".into()]);
+        assert_eq!(searcher.search(&from_regexes, &Count)?, 2);
+        assert_eq!(searcher.search(&from_patterns, &Count)?, 2);
+        Ok(())
+    }
+
     #[test]
     pub fn test_phrase_count() -> crate::Result<()> {
         let index = create_index(&["a c", "a a b d a b c", " a b"])?;
diff --git a/src/query/query_parser/query_parser.rs b/src/query/query_parser/query_parser.rs
index 73416ee2d..4d0185fdb 100644
--- a/src/query/query_parser/query_parser.rs
+++ b/src/query/query_parser/query_parser.rs
@@ -1106,6 +1106,7 @@ fn convert_to_query(fuzzy: &FxHashMap<Field, Fuzzy>, logical_ast: LogicalAst) ->
 #[cfg(test)]
 mod test {
     use matches::assert_matches;
+    use proptest::prelude::*;
 
     use super::super::logical_ast::*;
     use super::{QueryParser, QueryParserError};
@@ -1171,6 +1172,31 @@ mod test {
         make_query_parser_with_default_fields(&["title", "text"])
     }
 
+    proptest! {
+        #[test]
+        fn test_query_parser_does_not_panic_after_match_all(
+            suffix in prop::sample::select(vec!['\u{b}', '\u{c}', '\u{85}'])
+        ) {
+            let query_parser = make_query_parser();
+            let query = format!("*{suffix}");
+            prop_assert!(query_parser.parse_query(&query).is_err());
+            let (_, errors) = query_parser.parse_query_lenient(&query);
+            prop_assert!(!errors.is_empty());
+        }
+
+        #[test]
+        fn test_lenient_query_parser_makes_progress_in_invalid_sets(
+            field in proptest::option::of("[a-z]{1,4}"),
+            invalid_char in prop::sample::select(vec!['\0', '\u{b}', '\u{c}', '\u{7f}']),
+        ) {
+            let query_parser = make_query_parser();
+            let field = field.map(|field| format!("{field}:")).unwrap_or_default();
+            let query = format!("{field}IN [{invalid_char}");
+            let (_, errors) = query_parser.parse_query_lenient(&query);
+            prop_assert!(!errors.is_empty());
+        }
+    }
+
     fn parse_query_to_logical_ast_with_default_fields(
         query: &str,
         default_conjunction: bool,
diff --git a/src/snippet/mod.rs b/src/snippet/mod.rs
index ee61b534a..b097dacd8 100644
--- a/src/snippet/mod.rs
+++ b/src/snippet/mod.rs
@@ -115,6 +115,7 @@ impl FragmentCandidate {
 #[derive(Debug)]
 pub struct Snippet {
     fragment: String,
+    fragment_range: Range<usize>,
     highlighted: Vec<Range<usize>>,
     snippet_prefix: String,
     snippet_postfix: String,
@@ -122,9 +123,10 @@ pub struct Snippet {
 
 impl Snippet {
     /// Create a new `Snippet`.
-    fn new(fragment: &str, highlighted: Vec<Range<usize>>) -> Self {
+    fn new(fragment: &str, fragment_range: Range<usize>, highlighted: Vec<Range<usize>>) -> Self {
         Self {
             fragment: fragment.to_string(),
+            fragment_range,
             highlighted,
             snippet_prefix: DEFAULT_SNIPPET_PREFIX.to_string(),
             snippet_postfix: DEFAULT_SNIPPET_POSTFIX.to_string(),
@@ -135,6 +137,7 @@ impl Snippet {
     pub fn empty() -> Snippet {
         Snippet {
             fragment: String::new(),
+            fragment_range: 0..0,
             highlighted: Vec::new(),
             snippet_prefix: String::new(),
             snippet_postfix: String::new(),
@@ -169,6 +172,15 @@ impl Snippet {
         &self.fragment
     }
 
+    /// Returns the byte range of the fragment within the text the snippet was
+    /// generated from.
+    ///
+    /// For [`SnippetGenerator::snippet_from_doc`], that text is the field's
+    /// values joined by a single space and trimmed.
+    pub fn fragment_range(&self) -> Range<usize> {
+        self.fragment_range.clone()
+    }
+
     /// Returns a list of highlighted positions from the `Snippet`.
     pub fn highlighted(&self) -> &[Range<usize>] {
         &self.highlighted
@@ -250,7 +262,11 @@ fn select_best_fragment_combination(fragments: &[FragmentCandidate], text: &str)
             .iter()
             .map(|item| item.start - fragment.start_offset..item.end - fragment.start_offset)
             .collect();
-        Snippet::new(fragment_text, highlighted)
+        Snippet::new(
+            fragment_text,
+            fragment.start_offset..fragment.stop_offset,
+            highlighted,
+        )
     } else {
         // When there are no fragments to chose from,
         // for now create an empty snippet.
@@ -596,6 +612,7 @@ Survey in 2016, 2017, and 2018."#;
 
         let snippet = select_best_fragment_combination(&fragments[..], text);
         assert_eq!(snippet.fragment, "c d");
+        assert_eq!(snippet.fragment_range(), 4..7);
         assert_eq!(snippet.to_html(), "<b>c</b> d");
     }
 
@@ -619,9 +636,30 @@ Survey in 2016, 2017, and 2018."#;
 
         let snippet = select_best_fragment_combination(&fragments[..], text);
         assert_eq!(snippet.fragment, "e f");
+        assert_eq!(snippet.fragment_range(), 8..11);
         assert_eq!(snippet.to_html(), "e <b>f</b>");
     }
 
+    #[test]
+    fn test_snippet_fragment_range_with_repeated_text() {
+        // "a b" also occurs earlier, straddling a fragment boundary, so a
+        // text search for the fragment would locate the wrong occurrence.
+        let text = "x a b y a b";
+
+        let mut terms = BTreeMap::new();
+        terms.insert(String::from("a"), 1.0);
+        terms.insert(String::from("b"), 1.0);
+
+        let fragments =
+            search_fragments(&mut From::from(SimpleTokenizer::default()), text, &terms, 3);
+
+        let snippet = select_best_fragment_combination(&fragments[..], text);
+        assert_eq!(snippet.fragment, "a b");
+        assert_eq!(snippet.fragment_range(), 8..11);
+        assert_eq!(&text[snippet.fragment_range()], snippet.fragment());
+        assert_ne!(text.find(snippet.fragment()), Some(8));
+    }
+
     #[test]
     fn test_snippet_with_second_fragment_has_the_highest_score() {
         let text = "a b c d e f g";
@@ -675,6 +713,7 @@ Survey in 2016, 2017, and 2018."#;
 
         let snippet = select_best_fragment_combination(&fragments[..], text);
         assert_eq!(snippet.fragment, "");
+        assert_eq!(snippet.fragment_range(), 0..0);
         assert_eq!(snippet.to_html(), "");
         assert!(snippet.is_empty());
     }
diff --git a/sstable/benches/stream_bench.rs b/sstable/benches/stream_bench.rs
index 70dcdd8e3..f8235ceea 100644
--- a/sstable/benches/stream_bench.rs
+++ b/sstable/benches/stream_bench.rs
@@ -1,13 +1,68 @@
 use std::collections::BTreeSet;
+use std::hint::black_box;
 use std::io;
 
 use common::file_slice::FileSlice;
 use criterion::{Criterion, criterion_group, criterion_main};
 use rand::rngs::StdRng;
 use rand::{Rng, SeedableRng};
+use tantivy_fst::Automaton;
 use tantivy_sstable::{Dictionary, MonotonicU64SSTable};
 
 const CHARSET: &[u8] = b"abcdefghij";
+const AUTOMATON_PREFIX: &[u8] = b"ab";
+const NUM_AUTOMATON_MATCHES: usize = 1_017;
+
+// Matches `prefix.*`, but only implement can_match/will_always_match if configured to
+//
+// this allow comparing effects of optimisations depending on these functions
+struct HintedPrefixAutomaton<'a> {
+    prefix: &'a [u8],
+    can_match_hint: bool,
+    always_match_hint: bool,
+}
+
+impl<'a> HintedPrefixAutomaton<'a> {
+    fn new(prefix: &'a [u8], can_match_hint: bool, always_match_hint: bool) -> Self {
+        Self {
+            prefix,
+            can_match_hint,
+            always_match_hint,
+        }
+    }
+}
+
+impl Automaton for HintedPrefixAutomaton<'_> {
+    type State = Option<usize>;
+
+    fn start(&self) -> Self::State {
+        Some(0)
+    }
+
+    fn is_match(&self, state: &Self::State) -> bool {
+        *state == Some(self.prefix.len())
+    }
+
+    fn can_match(&self, state: &Self::State) -> bool {
+        !self.can_match_hint || state.is_some()
+    }
+
+    fn will_always_match(&self, state: &Self::State) -> bool {
+        self.always_match_hint && self.is_match(state)
+    }
+
+    fn accept(&self, state: &Self::State, byte: u8) -> Self::State {
+        let Some(pos) = *state else { return None };
+        if pos == self.prefix.len() {
+            return Some(pos);
+        }
+        if self.prefix[pos] == byte {
+            Some(pos + 1)
+        } else {
+            None
+        }
+    }
+}
 
 fn generate_key(rng: &mut impl Rng) -> String {
     let len = rng.random_range(3..12);
@@ -56,6 +111,26 @@ fn stream_bench(
     count
 }
 
+fn automaton_bench(
+    dictionary: &Dictionary<MonotonicU64SSTable>,
+    can_match_hint: bool,
+    always_match_hint: bool,
+) -> usize {
+    let mut stream = dictionary
+        .search(HintedPrefixAutomaton::new(
+            AUTOMATON_PREFIX,
+            black_box(can_match_hint),
+            black_box(always_match_hint),
+        ))
+        .into_stream()
+        .unwrap();
+    let mut count = 0;
+    while stream.advance() {
+        count += 1;
+    }
+    count
+}
+
 pub fn criterion_benchmark(c: &mut Criterion) {
     let dict = prepare_sstable().unwrap();
     c.bench_function("short_scan_init", |b| {
@@ -63,7 +138,7 @@ pub fn criterion_benchmark(c: &mut Criterion) {
     });
     c.bench_function("short_scan_init_and_scan", |b| {
         b.iter(|| {
-            assert_eq!(stream_bench(&dict, b"fa", b"faz", true), 971);
+            assert_eq!(stream_bench(&dict, b"fa", b"faz", true), 1051);
         })
     });
     c.bench_function("full_scan_init_and_scan_full_with_bound", |b| {
@@ -81,6 +156,18 @@ pub fn criterion_benchmark(c: &mut Criterion) {
             count
         })
     });
+    c.bench_function("full_scan_prefix_automaton_no_hints", |b| {
+        b.iter(|| assert_eq!(automaton_bench(&dict, false, false), NUM_AUTOMATON_MATCHES))
+    });
+    c.bench_function("full_scan_prefix_automaton_can_match_hint_only", |b| {
+        b.iter(|| assert_eq!(automaton_bench(&dict, true, false), NUM_AUTOMATON_MATCHES))
+    });
+    c.bench_function("full_scan_prefix_automaton_always_match_hint_only", |b| {
+        b.iter(|| assert_eq!(automaton_bench(&dict, false, true), NUM_AUTOMATON_MATCHES))
+    });
+    c.bench_function("full_scan_prefix_automaton_both_hints", |b| {
+        b.iter(|| assert_eq!(automaton_bench(&dict, true, true), NUM_AUTOMATON_MATCHES))
+    });
 }
 
 criterion_group!(benches, criterion_benchmark);
diff --git a/sstable/src/delta.rs b/sstable/src/delta.rs
index 97d868e4e..986700fef 100644
--- a/sstable/src/delta.rs
+++ b/sstable/src/delta.rs
@@ -1,3 +1,4 @@
+use std::cmp::Ordering;
 use std::io::{self, BufWriter, Write};
 use std::ops::Range;
 
@@ -12,6 +13,62 @@ const FOUR_BIT_LIMITS: usize = 1 << 4;
 const VINT_MODE: u8 = 1u8;
 const BLOCK_LEN: usize = 4_000;
 
+/// Incrementally compares delta-encoded keys with a fixed target key.
+pub(crate) struct DeltaKeyComparator {
+    num_matching_bytes: usize,
+}
+
+impl DeltaKeyComparator {
+    pub(crate) fn new() -> Self {
+        DeltaKeyComparator {
+            num_matching_bytes: 0,
+        }
+    }
+
+    #[inline(always)]
+    pub(crate) fn compare(
+        &mut self,
+        target: &[u8],
+        common_prefix_len: usize,
+        suffix: &[u8],
+    ) -> Ordering {
+        match common_prefix_len.cmp(&self.num_matching_bytes) {
+            // popped bytes already matched => too far
+            Ordering::Less => return Ordering::Greater,
+            Ordering::Equal => (),
+            // the ok prefix is less than current entry prefix => continue to next element
+            Ordering::Greater => return Ordering::Less,
+        }
+
+        for (key_byte, target_byte) in suffix.iter().zip(&target[self.num_matching_bytes..]) {
+            match key_byte.cmp(target_byte) {
+                Ordering::Equal => self.num_matching_bytes += 1,
+                ordering => return ordering,
+            }
+        }
+
+        (common_prefix_len + suffix.len()).cmp(&target.len())
+    }
+
+    #[inline(always)]
+    pub(crate) fn compare_across_blocks(
+        &mut self,
+        target: &[u8],
+        common_prefix_len: usize,
+        suffix: &[u8],
+    ) -> Ordering {
+        // blocks are independent. On each new block we get a common_prefix_len=0 entry.
+        // reset our state with it
+        if common_prefix_len == 0 {
+            let num_matching_bytes = crate::common_prefix_len(target, suffix);
+            self.num_matching_bytes = num_matching_bytes;
+            // cannot panicm at worth we might compare empty slices if num_matching_bytes==len()
+            return suffix[num_matching_bytes..].cmp(&target[num_matching_bytes..]);
+        }
+        self.compare(target, common_prefix_len, suffix)
+    }
+}
+
 pub struct DeltaWriter<W, TValueWriter>
 where W: io::Write
 {
@@ -241,7 +298,9 @@ where TValueReader: value::ValueReader
 
 #[cfg(test)]
 mod tests {
-    use super::DeltaReader;
+    use std::cmp::Ordering;
+
+    use super::{DeltaKeyComparator, DeltaReader};
     use crate::value::U64MonotonicValueReader;
 
     #[test]
@@ -249,4 +308,24 @@ mod tests {
         let mut delta_reader: DeltaReader<U64MonotonicValueReader> = DeltaReader::empty();
         assert!(!delta_reader.advance().unwrap());
     }
+
+    #[test]
+    fn test_delta_key_comparator_across_block_reset() {
+        let mut comparator = DeltaKeyComparator::new();
+        let target = b"bba";
+
+        assert_eq!(
+            comparator.compare_across_blocks(target, 0, b"baaaaa"),
+            Ordering::Less
+        );
+        assert_eq!(
+            comparator.compare_across_blocks(target, 2, b"baaa"),
+            Ordering::Less
+        );
+        // A zero-length common prefix marks a block reset, so the suffix is a complete key.
+        assert_eq!(
+            comparator.compare_across_blocks(target, 0, b"bbbaaa"),
+            Ordering::Greater
+        );
+    }
 }
diff --git a/sstable/src/dictionary.rs b/sstable/src/dictionary.rs
index 5de411467..69b57053f 100644
--- a/sstable/src/dictionary.rs
+++ b/sstable/src/dictionary.rs
@@ -14,6 +14,7 @@ use itertools::Itertools;
 use tantivy_fst::Automaton;
 use tantivy_fst::automaton::AlwaysMatch;
 
+use crate::delta::DeltaKeyComparator;
 use crate::streamer::{Streamer, StreamerBuilder};
 use crate::{BlockAddr, DeltaReader, Reader, SSTable, SSTableIndex, TermOrdinal, VoidSSTable};
 
@@ -356,41 +357,19 @@ impl<TSSTable: SSTable> Dictionary<TSSTable> {
     ) -> io::Result<TermOrdHit> {
         let mut term_ord = 0;
         let key_bytes = key.as_ref();
-        let mut ok_bytes = 0;
+        let mut key_comparator = DeltaKeyComparator::new();
         while sstable_delta_reader.advance()? {
-            let prefix_len = sstable_delta_reader.common_prefix_len();
-            let suffix = sstable_delta_reader.suffix();
-
-            match prefix_len.cmp(&ok_bytes) {
-                Ordering::Less => return Ok(TermOrdHit::Next(term_ord)), /* popped bytes already matched => too far */
-                Ordering::Equal => (),
-                Ordering::Greater => {
-                    // the ok prefix is less than current entry prefix => continue to next elem
+            match key_comparator.compare(
+                key_bytes,
+                sstable_delta_reader.common_prefix_len(),
+                sstable_delta_reader.suffix(),
+            ) {
+                Ordering::Less => {
                     term_ord += 1;
-                    continue;
                 }
+                Ordering::Equal => return Ok(TermOrdHit::Exact(term_ord)),
+                Ordering::Greater => return Ok(TermOrdHit::Next(term_ord)),
             }
-
-            // we have ok_bytes byte of common prefix, check if this key adds more
-            for (key_byte, suffix_byte) in key_bytes[ok_bytes..].iter().zip(suffix) {
-                match suffix_byte.cmp(key_byte) {
-                    Ordering::Less => break,          // byte too small
-                    Ordering::Equal => ok_bytes += 1, // new matching
-                    // byte
-                    Ordering::Greater => return Ok(TermOrdHit::Next(term_ord)), // too far
-                }
-            }
-
-            if ok_bytes == key_bytes.len() {
-                if prefix_len + suffix.len() == ok_bytes {
-                    return Ok(TermOrdHit::Exact(term_ord));
-                } else {
-                    // current key is a prefix of current element, not a match
-                    return Ok(TermOrdHit::Next(term_ord));
-                }
-            }
-
-            term_ord += 1;
         }
 
         Ok(TermOrdHit::Next(term_ord))
diff --git a/sstable/src/lib.rs b/sstable/src/lib.rs
index 1f6bd14c7..d925ad540 100644
--- a/sstable/src/lib.rs
+++ b/sstable/src/lib.rs
@@ -70,7 +70,7 @@ const SSTABLE_VERSION: u32 = 3;
 
 /// Given two byte string returns the length of
 /// the longest common prefix.
-fn common_prefix_len(left: &[u8], right: &[u8]) -> usize {
+pub(crate) fn common_prefix_len(left: &[u8], right: &[u8]) -> usize {
     left.iter()
         .cloned()
         .zip(right.iter().cloned())
diff --git a/sstable/src/streamer.rs b/sstable/src/streamer.rs
index 9203b3d0a..9e5d11541 100644
--- a/sstable/src/streamer.rs
+++ b/sstable/src/streamer.rs
@@ -1,9 +1,11 @@
+use std::cmp::Ordering;
 use std::io;
 use std::ops::Bound;
 
 use tantivy_fst::Automaton;
 use tantivy_fst::automaton::AlwaysMatch;
 
+use crate::delta::DeltaKeyComparator;
 use crate::dictionary::Dictionary;
 use crate::{DeltaReader, SSTable, TermOrdinal};
 
@@ -30,6 +32,22 @@ fn bound_as_byte_slice(bound: &Bound<Vec<u8>>) -> Bound<&[u8]> {
     }
 }
 
+#[inline(always)]
+fn matches_upper_bound(
+    comparator: &mut DeltaKeyComparator,
+    upper_bound: &Bound<Vec<u8>>,
+    common_prefix_len: usize,
+    suffix: &[u8],
+) -> bool {
+    let (upper_bound_key, inclusive) = match upper_bound {
+        Bound::Unbounded => return true,
+        Bound::Included(upper_bound_key) => (upper_bound_key, true),
+        Bound::Excluded(upper_bound_key) => (upper_bound_key, false),
+    };
+    let ordering = comparator.compare_across_blocks(upper_bound_key, common_prefix_len, suffix);
+    ordering == Ordering::Less || inclusive && ordering == Ordering::Equal
+}
+
 impl<'a, TSSTable, A> StreamerBuilder<'a, TSSTable, A>
 where
     A: Automaton,
@@ -124,14 +142,23 @@ where
             Bound::Unbounded => 0,
         };
 
+        let always_match_at = if self.automaton.will_always_match(&start_state) {
+            Some(0)
+        } else {
+            None
+        };
+
         Ok(Streamer {
             automaton: self.automaton,
             states: vec![start_state],
+            always_match_at,
             delta_reader,
             key: Vec::new(),
             term_ord: first_term.checked_sub(1),
+            lower_bound_reached: self.lower == Bound::Unbounded,
             lower_bound: self.lower,
             upper_bound: self.upper,
+            upper_bound_comparator: DeltaKeyComparator::new(),
             _lifetime: std::marker::PhantomData,
         })
     }
@@ -174,8 +201,11 @@ where
     term_ord: Option<TermOrdinal>,
     lower_bound: Bound<Vec<u8>>,
     upper_bound: Bound<Vec<u8>>,
+    upper_bound_comparator: DeltaKeyComparator,
     // this field is used to please the type-interface of a dictionary in tantivy
     _lifetime: std::marker::PhantomData<&'a ()>,
+    lower_bound_reached: bool,
+    always_match_at: Option<usize>,
 }
 
 impl<TSSTable> Streamer<'_, TSSTable, AlwaysMatch>
@@ -184,12 +214,15 @@ where TSSTable: SSTable
     pub fn empty() -> Self {
         Streamer {
             automaton: AlwaysMatch,
-            states: Vec::new(),
+            states: vec![AlwaysMatch.start()],
+            always_match_at: Some(0),
             delta_reader: DeltaReader::empty(),
             key: Vec::new(),
             term_ord: None,
+            lower_bound_reached: true,
             lower_bound: Bound::Unbounded,
             upper_bound: Bound::Unbounded,
+            upper_bound_comparator: DeltaKeyComparator::new(),
             _lifetime: std::marker::PhantomData,
         }
     }
@@ -201,53 +234,173 @@ where
     A::State: Clone,
     TSSTable: SSTable,
 {
+    #[inline(always)]
+    fn advance_delta_reader(&mut self) -> bool {
+        if !self.delta_reader.advance().unwrap() {
+            return false;
+        }
+        // An automaton prunes whole blocks, so the ordinal is not simply the previous one
+        // plus one: on entering a new slice it jumps to that slice's first term ordinal.
+        // Counting alone would report a term's position among the blocks actually scanned.
+        self.term_ord = Some(match self.delta_reader.take_first_ordinal() {
+            Some(first_ordinal) => first_ordinal,
+            None => self
+                .term_ord
+                .map(|term_ord| term_ord + 1u64)
+                .unwrap_or(0u64),
+        });
+        true
+    }
+
+    /// Make progress up to the lower bound
+    ///
+    /// Returns whether the reader was positioned on a key matching the lower bound.
+    /// If false, the delta_reader has been exhausted without finding such a key.
+    fn initialize(&mut self) -> bool {
+        debug_assert!(!self.lower_bound_reached);
+        let mut lower_bound_comparator = DeltaKeyComparator::new();
+        while self.advance_delta_reader() {
+            let common_prefix_len = self.delta_reader.common_prefix_len();
+            let suffix = self.delta_reader.suffix();
+            let (lower_bound_key, inclusive) = match &self.lower_bound {
+                Bound::Unbounded => unreachable!("unbounded streamers do not need initialization"),
+                Bound::Included(lower_bound_key) => (lower_bound_key, true),
+                Bound::Excluded(lower_bound_key) => (lower_bound_key, false),
+            };
+            let ordering = lower_bound_comparator.compare_across_blocks(
+                lower_bound_key,
+                common_prefix_len,
+                suffix,
+            );
+            let match_lower_bound =
+                ordering == Ordering::Greater || inclusive && ordering == Ordering::Equal;
+            if match_lower_bound {
+                self.key.clear();
+                self.key
+                    .extend_from_slice(&lower_bound_key[..common_prefix_len]);
+                self.key.extend_from_slice(suffix);
+                let mut state: A::State = self.states.last().unwrap().clone();
+                for b in &self.key {
+                    state = self.automaton.accept(&state, *b);
+                    self.states.push(state.clone());
+                }
+                self.lower_bound_reached = true;
+                return true;
+            }
+        }
+        self.lower_bound_reached = true;
+        false
+    }
+
     /// Advance position the stream on the next item.
     /// Before the first call to `.advance()`, the stream
     /// is an uninitialized state.
     pub fn advance(&mut self) -> bool {
-        while self.delta_reader.advance().unwrap() {
-            // An automaton prunes whole blocks, so the ordinal is not simply the previous one
-            // plus one: on entering a new slice it jumps to that slice's first term ordinal.
-            // Counting alone would report a term's position among the blocks actually scanned.
-            self.term_ord = Some(match self.delta_reader.take_first_ordinal() {
-                Some(first_ordinal) => first_ordinal,
-                None => self
-                    .term_ord
-                    .map(|term_ord| term_ord + 1u64)
-                    .unwrap_or(0u64),
-            });
+        if !self.lower_bound_reached {
+            if !self.initialize() {
+                // no key higher than lower-bound at all
+                return false;
+            }
+            if !matches_upper_bound(
+                &mut self.upper_bound_comparator,
+                &self.upper_bound,
+                0,
+                &self.key,
+            ) {
+                return false;
+            }
+            if self.automaton.is_match(self.states.last().unwrap()) {
+                return true;
+            }
+        }
+
+        match (
+            // we could check always_match_at == Some(0), but this actually gets
+            // inlined into `true` with AlwaysMatch, which is even faster
+            self.automaton
+                .will_always_match(&self.states.first().unwrap()),
+            self.upper_bound == Bound::Unbounded,
+        ) {
+            (true, true) => self.advance_always_match::<true>(),
+            (true, false) => self.advance_always_match::<false>(),
+            (false, true) => self.advance_with_automaton::<true>(),
+            (false, false) => self.advance_with_automaton::<false>(),
+        }
+    }
+
+    fn advance_always_match<const NO_BOUND: bool>(&mut self) -> bool {
+        if !self.advance_delta_reader() {
+            return false;
+        }
+        self.reconstruct_key_and_check_upper_bound::<NO_BOUND>()
+    }
+
+    fn advance_with_automaton<const NO_BOUND: bool>(&mut self) -> bool {
+        // fast path, check if prefix always match and we can skip Vec<state> management
+        if let Some(always_match_at) = self.always_match_at.take() {
+            if !self.advance_delta_reader() {
+                return false;
+            }
+            let common_prefix_len = self.delta_reader.common_prefix_len();
+            if always_match_at <= common_prefix_len {
+                self.always_match_at = Some(always_match_at);
+                return self.reconstruct_key_and_check_upper_bound::<NO_BOUND>();
+            }
+        } else if !self.advance_delta_reader() {
+            return false;
+        }
+
+        loop {
             let common_prefix_len = self.delta_reader.common_prefix_len();
             self.states.truncate(common_prefix_len + 1);
-            self.key.truncate(common_prefix_len);
+            // TODO we could detect when we reach a !can_match, and skip both state and key
+            // computation until we truncate that can_t_match out of our state. it's already
+            // done at the block layer, so not as important
             let mut state: A::State = self.states.last().unwrap().clone();
             for &b in self.delta_reader.suffix() {
                 state = self.automaton.accept(&state, b);
                 self.states.push(state.clone());
             }
-            self.key.extend_from_slice(self.delta_reader.suffix());
-            let match_lower_bound = match &self.lower_bound {
-                Bound::Unbounded => true,
-                Bound::Included(lower_bound_key) => lower_bound_key[..] <= self.key[..],
-                Bound::Excluded(lower_bound_key) => lower_bound_key[..] < self.key[..],
-            };
-            if !match_lower_bound {
-                continue;
+            let matches = self.automaton.is_match(&state);
+            if matches {
+                self.always_match_at = self
+                    .states
+                    .iter()
+                    .enumerate()
+                    .rev()
+                    .take_while(|(_i, state)| self.automaton.will_always_match(state))
+                    .last()
+                    .map(|(i, _state)| i);
             }
-            // We match the lower key once. All subsequent keys will pass that bar.
-            self.lower_bound = Bound::Unbounded;
-            let match_upper_bound = match &self.upper_bound {
-                Bound::Unbounded => true,
-                Bound::Included(upper_bound_key) => upper_bound_key[..] >= self.key[..],
-                Bound::Excluded(upper_bound_key) => upper_bound_key[..] > self.key[..],
-            };
-            if !match_upper_bound {
+
+            if !self.reconstruct_key_and_check_upper_bound::<NO_BOUND>() {
                 return false;
             }
-            if self.automaton.is_match(&state) {
+            if matches {
                 return true;
             }
+            if !self.advance_delta_reader() {
+                return false;
+            }
         }
-        false
+    }
+
+    #[inline(always)]
+    fn reconstruct_key_and_check_upper_bound<const NO_BOUND: bool>(&mut self) -> bool {
+        let common_prefix_len = self.delta_reader.common_prefix_len();
+        self.key.truncate(common_prefix_len);
+        self.key.extend_from_slice(self.delta_reader.suffix());
+
+        // TODO there is an idea where we only look at the upper bound when our delta_reader
+        // reached the last block (if we pruned blocks beforehand (do we always?) we cannot
+        // find that key before that block)
+        NO_BOUND
+            || matches_upper_bound(
+                &mut self.upper_bound_comparator,
+                &self.upper_bound,
+                common_prefix_len,
+                self.delta_reader.suffix(),
+            )
     }
 
     /// Returns the `TermOrdinal` of the given term.