docs(node): regenerate RRF reranker reference

docs/src/js/namespaces/rerankers/classes/RRFReranker.md

@@ -10,6 +10,24 @@ Reranks the results using the Reciprocal Rank Fusion (RRF) algorithm.
 
 ## Methods
 
+### outputSchema()
+
+```ts
+outputSchema(inputSchema): Promise<Schema<any>>
+```
+
+Declare the RRF output schema for vector-only query execution.
+
+#### Parameters
+
+* **inputSchema**: `Schema`&lt;`any`&gt;
+
+#### Returns
+
+`Promise`&lt;`Schema`&lt;`any`&gt;&gt;
+
+***
+
 ### rerankHybrid()
 
 ```ts

docs/src/js/namespaces/rerankers/interfaces/Reranker.md

@@ -8,6 +8,27 @@
 
 ## Methods
 
+### outputSchema()?
+
+```ts
+optional outputSchema(inputSchema): Promise<Schema<any>>
+```
+
+Declare the schema returned when reranking a vector-only query.
+
+This is required for vector-only reranking so query schema introspection
+and execution agree. Hybrid-only rerankers may omit it.
+
+#### Parameters
+
+* **inputSchema**: `Schema`&lt;`any`&gt;
+
+#### Returns
+
+`Promise`&lt;`Schema`&lt;`any`&gt;&gt;
+
+***
+
 ### rerankHybrid()
 
 ```ts

nodejs/__test__/rerankers.test.ts

@@ -79,6 +79,22 @@ describe("rerankers", function () {
     expect(result).toHaveLength(2);
   });
 
+  it("returns relevance scores when reranking a vector search", async function () {
+    const query = table
+      .vectorSearch([0.1, 0.1])
+      .limit(2)
+      .rerank(await RRFReranker.create());
+    const schema = await query.outputSchema();
+    const result = await query.toArray();
+
+    expect(schema.fields.map((field) => field.name)).toContain(
+      "_relevance_score",
+    );
+    expect(result).toHaveLength(2);
+    expect(result[0]._relevance_score).toBeCloseTo(1 / 60);
+    expect(result[1]._relevance_score).toBeCloseTo(1 / 61);
+  });
+
   it("does not keep process alive after rerank query", async function () {
     const script = `
 import * as lancedb from "./dist/index.js";

nodejs/lancedb/query.ts

@@ -5,9 +5,11 @@ import {
   Table as ArrowTable,
   type IntoVector,
   RecordBatch,
+  createEmptyTable,
   extractVectorBuffer,
   fromBufferToRecordBatch,
   fromRecordBatchToBuffer,
+  fromTableToBuffer,
   tableFromIPC,
 } from "./arrow";
 import { type IvfPqOptions } from "./indices";
@@ -744,20 +746,30 @@ export class VectorQuery extends StandardQueryBase<NativeVectorQuery> {
   }
 
   rerank(reranker: Reranker): VectorQuery {
-    super.doCall((inner) =>
-      inner.rerank(async (args) => {
-        const vecResults = await fromBufferToRecordBatch(args.vecResults);
-        const ftsResults = await fromBufferToRecordBatch(args.ftsResults);
-        const result = await reranker.rerankHybrid(
-          args.query,
-          vecResults as RecordBatch,
-          ftsResults as RecordBatch,
-        );
+    super.doCall((inner) => {
+      const outputSchema = reranker.outputSchema?.bind(reranker);
+      inner.rerank(
+        async (args) => {
+          const vecResults = await fromBufferToRecordBatch(args.vecResults);
+          const ftsResults = await fromBufferToRecordBatch(args.ftsResults);
+          const result = await reranker.rerankHybrid(
+            args.query,
+            vecResults as RecordBatch,
+            ftsResults as RecordBatch,
+          );
 
-        const buffer = fromRecordBatchToBuffer(result);
-        return buffer;
-      }),
-    );
+          const buffer = fromRecordBatchToBuffer(result);
+          return buffer;
+        },
+        outputSchema
+          ? async (args) => {
+              const inputSchema = tableFromIPC(args.inputSchema).schema;
+              const result = await outputSchema(inputSchema);
+              return fromTableToBuffer(createEmptyTable(result));
+            }
+          : undefined,
+      );
+    });
 
     return this;
   }

nodejs/lancedb/rerankers/index.ts

@@ -1,14 +1,22 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
-import { RecordBatch } from "apache-arrow";
+import { RecordBatch, Schema } from "apache-arrow";
 
 export * from "./rrf";
 
-// Interface for a reranker. A reranker is used to rerank the results from a
-// vector and FTS search. This is useful for combining the results from both
-// search methods.
+// Interface for a reranker. A reranker is used to rerank vector and hybrid
+// search results. For vector-only searches, query is empty and ftsResults is an
+// empty batch with the same schema as vecResults.
 export interface Reranker {
+  /**
+   * Declare the schema returned when reranking a vector-only query.
+   *
+   * This is required for vector-only reranking so query schema introspection
+   * and execution agree. Hybrid-only rerankers may omit it.
+   */
+  outputSchema?(inputSchema: Schema): Promise<Schema>;
+
   rerankHybrid(
     query: string,
     vecResults: RecordBatch,

nodejs/lancedb/rerankers/rrf.ts

@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
-import { RecordBatch } from "apache-arrow";
+import { Field, Float32, RecordBatch, Schema } from "apache-arrow";
 import { fromBufferToRecordBatch, fromRecordBatchToBuffer } from "../arrow";
 import { RrfReranker as NativeRRFReranker } from "../native";
 
@@ -24,6 +24,17 @@ export class RRFReranker {
     );
   }
 
+  /** Declare the RRF output schema for vector-only query execution. */
+  async outputSchema(inputSchema: Schema): Promise<Schema> {
+    return new Schema(
+      [
+        ...inputSchema.fields,
+        new Field("_relevance_score", new Float32(), false),
+      ],
+      inputSchema.metadata,
+    );
+  }
+
   async rerankHybrid(
     query: string,
     vecResults: RecordBatch,

nodejs/src/query.rs

@@ -6,8 +6,8 @@ use std::sync::Arc;
 use crate::error::NapiErrorExt;
 use crate::error::convert_error;
 use crate::iterator::RecordBatchIterator;
-use crate::rerankers::RerankHybridCallbackArgs;
 use crate::rerankers::Reranker;
+use crate::rerankers::{RerankHybridCallbackArgs, RerankOutputSchemaCallbackArgs};
 use crate::util::{parse_distance_type, schema_to_buffer};
 use arrow_array::{
     Array, Float16Array as ArrowFloat16Array, Float32Array as ArrowFloat32Array,
@@ -388,8 +388,9 @@ impl VectorQuery {
     pub fn rerank(
         &mut self,
         rerank_hybrid: Function<RerankHybridCallbackArgs, Promise<Buffer>>,
+        output_schema: Option<Function<RerankOutputSchemaCallbackArgs, Promise<Buffer>>>,
     ) -> napi::Result<()> {
-        let reranker = Reranker::new(rerank_hybrid)?;
+        let reranker = Reranker::new(rerank_hybrid, output_schema)?;
         self.inner = self.inner.clone().rerank(Arc::new(reranker));
         Ok(())
     }

nodejs/src/rerankers.rs

@@ -6,7 +6,7 @@ use async_trait::async_trait;
 use napi::{bindgen_prelude::*, threadsafe_function::ThreadsafeFunction};
 use napi_derive::napi;
 
-use lancedb::ipc::batches_to_ipc_file;
+use lancedb::ipc::{batches_to_ipc_file, ipc_file_to_schema, schema_to_ipc_file};
 use lancedb::rerankers::Reranker as LanceDBReranker;
 use lancedb::{error::Error, ipc::ipc_file_to_batches};
 
@@ -21,28 +21,72 @@ type RerankHybridFn = ThreadsafeFunction<
     true,
 >;
 
+type RerankOutputSchemaFn = ThreadsafeFunction<
+    RerankOutputSchemaCallbackArgs,
+    Promise<Buffer>,
+    RerankOutputSchemaCallbackArgs,
+    Status,
+    false,
+    true,
+>;
+
 /// Reranker implementation that "wraps" a NodeJS Reranker implementation.
 /// This contains references to the callbacks that can be used to invoke the
 /// reranking methods on the NodeJS implementation and handles serializing the
 /// record batches to Arrow IPC buffers.
 pub struct Reranker {
     rerank_hybrid: RerankHybridFn,
+    output_schema: Option<RerankOutputSchemaFn>,
 }
 
 impl Reranker {
     pub fn new(
         rerank_hybrid: Function<RerankHybridCallbackArgs, Promise<Buffer>>,
+        output_schema: Option<Function<RerankOutputSchemaCallbackArgs, Promise<Buffer>>>,
     ) -> napi::Result<Self> {
         let rerank_hybrid = rerank_hybrid
             .build_threadsafe_function()
             .weak::<true>()
             .build()?;
-        Ok(Self { rerank_hybrid })
+        let output_schema = output_schema
+            .map(|output_schema| {
+                output_schema
+                    .build_threadsafe_function()
+                    .weak::<true>()
+                    .build()
+            })
+            .transpose()?;
+        Ok(Self {
+            rerank_hybrid,
+            output_schema,
+        })
     }
 }
 
 #[async_trait]
 impl lancedb::rerankers::Reranker for Reranker {
+    async fn output_schema(
+        &self,
+        input: &arrow_schema::SchemaRef,
+    ) -> lancedb::error::Result<arrow_schema::SchemaRef> {
+        let output_schema = self.output_schema.as_ref().ok_or(Error::NotSupported {
+            message: "vector rerankers must declare their output schema".to_string(),
+        })?;
+        let callback_args = RerankOutputSchemaCallbackArgs {
+            input_schema: Buffer::from(schema_to_ipc_file(input.as_ref())?),
+        };
+        let promised_buffer: Promise<Buffer> = output_schema
+            .call_async(callback_args)
+            .await
+            .map_err(|e| Error::Runtime {
+                message: format!("napi error status={}, reason={}", e.status, e.reason),
+            })?;
+        let buffer = promised_buffer.await.map_err(|e| Error::Runtime {
+            message: format!("napi error status={}, reason={}", e.status, e.reason),
+        })?;
+        ipc_file_to_schema(buffer.to_vec())
+    }
+
     async fn rerank_hybrid(
         &self,
         query: &str,
@@ -86,6 +130,11 @@ pub struct RerankHybridCallbackArgs {
     pub fts_results: Buffer,
 }
 
+#[napi(object)]
+pub struct RerankOutputSchemaCallbackArgs {
+    pub input_schema: Buffer,
+}
+
 fn buffer_to_record_batch(buffer: Buffer) -> Result<RecordBatch> {
     let mut reader = ipc_file_to_batches(buffer.to_vec()).default_error()?;
     reader

python/python/lancedb/embeddings/instructor.py

@@ -3,7 +3,6 @@
 
 
 from typing import List
-from urllib.parse import unquote, urlparse
 
 import numpy as np
 
@@ -126,20 +125,9 @@ class InstructorEmbeddingFunction(TextEmbeddingFunction):
 
     @weak_lru(maxsize=1)
     def get_model(self):
-        huggingface_hub = attempt_import_or_raise("huggingface_hub", "huggingface-hub")
-        missing = object()
-        original_cached_download = getattr(huggingface_hub, "cached_download", missing)
-        if original_cached_download is missing:
-            huggingface_hub.cached_download = _cached_download(huggingface_hub)
-
-        try:
-            instructor_embedding = attempt_import_or_raise(
-                "InstructorEmbedding", "InstructorEmbedding"
-            )
-        finally:
-            if original_cached_download is missing:
-                del huggingface_hub.cached_download
-
+        instructor_embedding = attempt_import_or_raise(
+            "InstructorEmbedding", "InstructorEmbedding"
+        )
         torch = attempt_import_or_raise("torch", "torch")
 
         model = instructor_embedding.INSTRUCTOR(self.name)
@@ -152,44 +140,3 @@ class InstructorEmbeddingFunction(TextEmbeddingFunction):
                 model, {torch.nn.Linear}, dtype=torch.qint8
             )
         return model
-
-
-def _cached_download(huggingface_hub):
-    """Provide the legacy download API used by sentence-transformers 2.2.x."""
-
-    def cached_download(
-        *,
-        url,
-        cache_dir=None,
-        force_filename=None,
-        library_name=None,
-        library_version=None,
-        user_agent=None,
-        use_auth_token=None,
-        **_,
-    ):
-        path = urlparse(url).path.lstrip("/")
-        try:
-            repo_id, resolved_path = path.split("/resolve/", maxsplit=1)
-            revision, filename = resolved_path.split("/", maxsplit=1)
-        except ValueError as err:
-            raise ValueError(f"Unsupported Hugging Face Hub URL: {url}") from err
-
-        repo_id = unquote(repo_id)
-        revision = unquote(revision)
-        filename = unquote(filename)
-        # sentence-transformers derives force_filename from this Hub path with
-        # os.path.join. Using the URL path beneath local_dir produces the same
-        # local destination without sending Windows separators to the Hub.
-        return huggingface_hub.hf_hub_download(
-            repo_id=repo_id,
-            filename=filename,
-            revision=revision,
-            local_dir=cache_dir,
-            library_name=library_name,
-            library_version=library_version,
-            user_agent=user_agent,
-            token=use_auth_token,
-        )
-
-    return cached_download

python/python/tests/test_embeddings.py

@@ -1,11 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
-import ntpath
 import os
 import pickle
-import sys
-from types import ModuleType
 from typing import List, Optional, Union
 from unittest.mock import MagicMock, patch
 
@@ -525,59 +522,6 @@ def test_embedding_function_safe_model_dump(embedding_type):
             )
 
 
-def test_instructor_embedding_supports_huggingface_hub_without_cached_download(
-    tmp_path, monkeypatch
-):
-    from lancedb.embeddings.instructor import InstructorEmbeddingFunction
-
-    hub_download = MagicMock(return_value="/cache/1_Pooling/config.json")
-    huggingface_hub = ModuleType("huggingface_hub")
-    huggingface_hub.hf_hub_download = hub_download
-    torch = ModuleType("torch")
-    monkeypatch.setitem(sys.modules, "huggingface_hub", huggingface_hub)
-    monkeypatch.setitem(sys.modules, "torch", torch)
-    monkeypatch.delitem(sys.modules, "InstructorEmbedding", raising=False)
-    monkeypatch.syspath_prepend(str(tmp_path))
-
-    (tmp_path / "InstructorEmbedding.py").write_text(
-        "from huggingface_hub import cached_download\n\n"
-        "class INSTRUCTOR:\n"
-        "    def __init__(self, name):\n"
-        "        self.name = name\n"
-    )
-
-    embedding = InstructorEmbeddingFunction.create(show_progress_bar=False)
-    instructor_model = embedding.get_model()
-
-    assert instructor_model.name == "hkunlp/instructor-base"
-    assert not hasattr(huggingface_hub, "cached_download")
-
-    instructor_embedding = sys.modules["InstructorEmbedding"]
-    path = instructor_embedding.cached_download(
-        url=(
-            "https://huggingface.co/hkunlp/instructor-base/resolve/abc123/"
-            "1_Pooling/config.json"
-        ),
-        cache_dir="/cache",
-        force_filename=ntpath.join("1_Pooling", "config.json"),
-        library_name="sentence-transformers",
-        library_version="2.2.2",
-        use_auth_token="token",
-    )
-
-    assert path == "/cache/1_Pooling/config.json"
-    hub_download.assert_called_once_with(
-        repo_id="hkunlp/instructor-base",
-        filename="1_Pooling/config.json",
-        revision="abc123",
-        local_dir="/cache",
-        library_name="sentence-transformers",
-        library_version="2.2.2",
-        user_agent=None,
-        token="token",
-    )
-
-
 @patch("time.sleep")
 def test_retry(mock_sleep):
     test_function = MagicMock(side_effect=[Exception] * 9 + ["result"])

rust/lancedb/src/query.rs

@@ -511,7 +511,9 @@ pub trait QueryBase {
 
     /// Rerank the results using the specified reranker.
     ///
-    /// This is currently only supported for Hybrid Search.
+    /// For vector-only searches, the reranker receives the vector results and an
+    /// empty full-text result set and must declare its output schema. Reranking
+    /// multiple query vectors in one query is not supported.
     fn rerank(self, reranker: Arc<dyn Reranker>) -> Self;
 
     /// The method to normalize the scores. Can be "rank" or "Score". If "Rank",
@@ -1138,6 +1140,44 @@ pub struct VectorQuery {
 }
 
 impl VectorQuery {
+    fn check_vector_rerank_supported(&self) -> Result<()> {
+        if self.request.query_vector.len() > 1 {
+            return Err(Error::NotSupported {
+                message: "reranking multiple query vectors is not supported; execute one query per vector"
+                    .to_string(),
+            });
+        }
+        Ok(())
+    }
+
+    async fn vector_rerank_output_schema(&self) -> Result<SchemaRef> {
+        self.check_vector_rerank_supported()?;
+
+        // Rerankers receive row IDs internally. Apply their schema transform to
+        // that exact input and then hide the row ID from the declared public
+        // schema unless it was explicitly requested.
+        let vector_query = self.clone().with_row_id();
+        let plan = vector_query
+            .create_plan(QueryExecutionOptions::default())
+            .await?;
+        let reranker = self
+            .request
+            .base
+            .reranker
+            .as_ref()
+            .expect("vector_rerank_output_schema requires a reranker");
+        let input_schema = plan.schema();
+        let output_schema = reranker.output_schema(&input_schema).await?;
+
+        if self.request.base.with_row_id {
+            Ok(output_schema)
+        } else {
+            Ok(RecordBatch::new_empty(output_schema)
+                .drop_column(ROW_ID)?
+                .schema())
+        }
+    }
+
     fn new(base: Query) -> Self {
         Self {
             parent: base.parent,
@@ -1443,6 +1483,61 @@ impl VectorQuery {
         Ok(single_batch_stream(results, max_batch_length))
     }
 
+    async fn execute_vector_rerank(
+        &self,
+        options: QueryExecutionOptions,
+    ) -> Result<SendableRecordBatchStream> {
+        self.check_vector_rerank_supported()?;
+
+        let max_batch_length = options.max_batch_length as usize;
+        let internal_options = options.without_output_batch_length_limit();
+
+        // RRF needs row IDs to assign and preserve scores. Keep them internal unless
+        // the caller explicitly requested them.
+        let vector_query = self.clone().with_row_id();
+        let vector_results = vector_query
+            .inner_execute_with_options(internal_options)
+            .await?;
+        let schema = vector_results.schema();
+        let vector_results = vector_results.try_collect::<Vec<_>>().await?;
+        let vector_results = concat_batches(&schema, vector_results.iter())?;
+        let vector_schema = vector_results.schema();
+        let fts_results = RecordBatch::new_empty(vector_schema.clone());
+
+        let reranker = self
+            .request
+            .base
+            .reranker
+            .as_ref()
+            .expect("execute_vector_rerank requires a reranker");
+        let expected_schema = reranker.output_schema(&vector_schema).await?;
+        let mut results = reranker
+            .rerank_hybrid("", vector_results, fts_results)
+            .await?;
+
+        check_reranker_result(&results)?;
+        if results.schema() != expected_schema {
+            return Err(Error::Schema {
+                message: format!(
+                    "reranker returned schema {:?}, but declared {:?}",
+                    results.schema(),
+                    expected_schema
+                ),
+            });
+        }
+
+        let limit = self.request.base.limit.unwrap_or(DEFAULT_TOP_K);
+        if results.num_rows() > limit {
+            results = results.slice(0, limit);
+        }
+
+        if !self.request.base.with_row_id {
+            results = results.drop_column(ROW_ID)?;
+        }
+
+        Ok(single_batch_stream(results, max_batch_length))
+    }
+
     async fn inner_execute_with_options(
         &self,
         options: QueryExecutionOptions,
@@ -1495,6 +1590,23 @@ impl ExecutableQuery for VectorQuery {
             return Ok(hybrid_result);
         }
 
+        if self.request.base.reranker.is_some() {
+            let timeout = options.timeout;
+            let mut rerank_options = options;
+            // A single outer deadline covers planning, candidate collection,
+            // schema declaration, and the complete reranker callback.
+            rerank_options.timeout = None;
+            let execution = self.execute_vector_rerank(rerank_options);
+            return match timeout {
+                Some(timeout) => tokio::time::timeout(timeout, execution)
+                    .await
+                    .map_err(|_| Error::Timeout {
+                        message: format!("Query timeout after {} ms", timeout.as_millis()),
+                    })?,
+                None => execution.await,
+            };
+        }
+
         self.inner_execute_with_options(options).await
     }
 
@@ -1507,6 +1619,15 @@ impl ExecutableQuery for VectorQuery {
         let query = AnyQuery::VectorQuery(self.request.clone());
         self.parent.analyze_plan(&query, options).await
     }
+
+    async fn output_schema(&self) -> Result<SchemaRef> {
+        if self.request.base.full_text_search.is_none() && self.request.base.reranker.is_some() {
+            self.vector_rerank_output_schema().await
+        } else {
+            let plan = self.create_plan(QueryExecutionOptions::default()).await?;
+            Ok(plan.schema())
+        }
+    }
 }
 
 impl HasQuery for VectorQuery {
@@ -1643,7 +1764,13 @@ impl ExecutableQuery for TakeQuery {
 
 #[cfg(test)]
 mod tests {
-    use std::{collections::HashSet, sync::Arc};
+    use std::{
+        collections::HashSet,
+        sync::{
+            Arc,
+            atomic::{AtomicBool, Ordering},
+        },
+    };
 
     use super::*;
     use arrow::{array::downcast_array, compute::concat_batches, datatypes::Int32Type};
@@ -1659,6 +1786,31 @@ mod tests {
 
     use crate::{Table, connect, database::CreateTableMode, index::Index};
 
+    #[derive(Debug)]
+    struct SlowReranker {
+        invoked: Arc<AtomicBool>,
+    }
+
+    #[async_trait::async_trait]
+    impl Reranker for SlowReranker {
+        async fn output_schema(&self, input: &SchemaRef) -> Result<SchemaRef> {
+            RRFReranker::default().output_schema(input).await
+        }
+
+        async fn rerank_hybrid(
+            &self,
+            query: &str,
+            vector_results: RecordBatch,
+            fts_results: RecordBatch,
+        ) -> Result<RecordBatch> {
+            self.invoked.store(true, Ordering::SeqCst);
+            tokio::time::sleep(Duration::from_secs(2)).await;
+            RRFReranker::default()
+                .rerank_hybrid(query, vector_results, fts_results)
+                .await
+        }
+    }
+
     #[tokio::test]
     async fn test_setters_getters() {
         // TODO: Switch back to memory://foo after https://github.com/lancedb/lancedb/issues/1051
@@ -2370,6 +2522,71 @@ mod tests {
         // We don't guarantee order.
         assert!(query_index.values().contains(&0));
         assert!(query_index.values().contains(&1));
+
+        let reranked = query.rerank(Arc::new(RRFReranker::default()));
+        let Err(execute_error) = reranked.execute().await else {
+            panic!("multi-vector reranking should be rejected");
+        };
+        assert!(
+            execute_error
+                .to_string()
+                .contains("reranking multiple query vectors is not supported")
+        );
+        let schema_error = reranked.output_schema().await.unwrap_err();
+        assert!(
+            schema_error
+                .to_string()
+                .contains("reranking multiple query vectors is not supported")
+        );
+    }
+
+    #[tokio::test]
+    async fn test_vector_rerank_timeout_covers_reranker() {
+        let tmp_dir = tempdir().unwrap();
+        let table = make_test_table(&tmp_dir).await;
+        let invoked = Arc::new(AtomicBool::new(false));
+        let reranker = SlowReranker {
+            invoked: invoked.clone(),
+        };
+
+        let result = table
+            .vector_search(&[0.1, 0.2, 0.3, 0.4])
+            .unwrap()
+            .limit(1)
+            .rerank(Arc::new(reranker))
+            .execute_with_options(QueryExecutionOptions {
+                timeout: Some(Duration::from_secs(1)),
+                ..Default::default()
+            })
+            .await;
+
+        assert!(invoked.load(Ordering::SeqCst));
+        assert!(matches!(result, Err(Error::Timeout { .. })));
+    }
+
+    #[tokio::test]
+    async fn test_vector_rerank_output_schema_matches_execution() {
+        let tmp_dir = tempdir().unwrap();
+        let table = make_test_table(&tmp_dir).await;
+        let query = table
+            .vector_search(&[0.1, 0.2, 0.3, 0.4])
+            .unwrap()
+            .limit(1)
+            .rerank(Arc::new(RRFReranker::default()));
+
+        let promised = query.output_schema().await.unwrap();
+        let actual = query
+            .execute()
+            .await
+            .unwrap()
+            .next()
+            .await
+            .unwrap()
+            .unwrap()
+            .schema();
+
+        assert_eq!(promised, actual);
+        assert!(promised.column_with_name("_relevance_score").is_some());
     }
 
     #[tokio::test]

rust/lancedb/src/rerankers.rs

@@ -8,6 +8,7 @@ use arrow::{
     compute::{concat_batches, filter_record_batch},
 };
 use arrow_array::{BooleanArray, RecordBatch, UInt64Array};
+use arrow_schema::SchemaRef;
 use async_trait::async_trait;
 use lance::dataset::ROW_ID;
 
@@ -47,16 +48,28 @@ impl std::fmt::Display for NormalizeMethod {
     }
 }
 
-/// Interface for a reranker. A reranker is used to rerank the results from a
-/// vector and FTS search. This is useful for combining the results from both
-/// search methods.
+/// Interface for a reranker. A reranker is used to rerank vector and hybrid
+/// search results. This is useful for combining results from multiple search
+/// methods or assigning a relevance score to vector search results.
 #[async_trait]
 pub trait Reranker: std::fmt::Debug + Sync + Send {
-    // TODO support vector reranking and FTS reranking. Currently only hybrid reranking is supported.
+    /// Declare the schema returned by [`Self::rerank_hybrid`] for a vector-only
+    /// query.
+    ///
+    /// Vector reranking validates the returned batch against this schema so
+    /// [`crate::query::ExecutableQuery::output_schema`] and execution cannot
+    /// disagree. Rerankers that only support hybrid search do not need to
+    /// implement this method.
+    async fn output_schema(&self, _input: &SchemaRef) -> Result<SchemaRef> {
+        Err(Error::NotSupported {
+            message: "vector rerankers must declare their output schema".to_string(),
+        })
+    }
 
     /// Rerank function receives the individual results from the vector and FTS search
     /// results. You can choose to use any of the results to generate the final results,
-    /// allowing maximum flexibility.
+    /// allowing maximum flexibility. For a vector-only search, `query` is empty and
+    /// `fts_results` is an empty batch with the same schema as `vector_results`.
     async fn rerank_hybrid(
         &self,
         query: &str,

rust/lancedb/src/rerankers/rrf.rs

@@ -9,7 +9,7 @@ use arrow::{
     compute::{sort_to_indices, take},
 };
 use arrow_array::{Float32Array, RecordBatch, UInt64Array};
-use arrow_schema::{DataType, Field, Schema, SortOptions};
+use arrow_schema::{DataType, Field, Schema, SchemaRef, SortOptions};
 use async_trait::async_trait;
 use lance::dataset::ROW_ID;
 
@@ -44,6 +44,19 @@ impl Default for RRFReranker {
 
 #[async_trait]
 impl Reranker for RRFReranker {
+    async fn output_schema(&self, input: &SchemaRef) -> Result<SchemaRef> {
+        let mut fields = input.fields().to_vec();
+        fields.push(Arc::new(Field::new(
+            RELEVANCE_SCORE,
+            DataType::Float32,
+            false,
+        )));
+        Ok(Arc::new(Schema::new_with_metadata(
+            fields,
+            input.metadata().clone(),
+        )))
+    }
+
     async fn rerank_hybrid(
         &self,
         _query: &str,
@@ -135,15 +148,9 @@ impl Reranker for RRFReranker {
             .collect();
 
         // add relevance score to schema
-        let mut fields = combined_results.schema().fields().to_vec();
-        fields.push(Arc::new(Field::new(
-            RELEVANCE_SCORE,
-            DataType::Float32,
-            false,
-        )));
-        let schema = Schema::new(fields);
+        let schema = self.output_schema(&combined_results.schema()).await?;
 
-        let combined_results = RecordBatch::try_new(Arc::new(schema), columns)?;
+        let combined_results = RecordBatch::try_new(schema, columns)?;
 
         Ok(combined_results)
     }