Merge main into gatekeeper/fix-3530-1

## Summary

- enrich local query field-not-found errors with recursively qualified
Arrow struct leaf paths
- preserve all other Lance and DataFusion errors unchanged
- add a regression test for the Python-visible filter error described in
the issue

## Root cause

DataFusion builds `FieldNotFound` candidates from the top-level Arrow
schema even though Lance supports dotted struct-field filters. As a
result, the error listed only the struct container and hid its valid
nested leaves.

## Validation

- `cargo test --quiet --features remote -p lancedb
table::query::tests::test_missing_filter_field_lists_nested_fields --
--exact`
- `cargo check --quiet --features remote --tests --examples`
- `cargo clippy --quiet --features remote --tests --examples` (passes
with pre-existing unrelated warnings)
- `cargo fmt --all -- --check`

Fixes #951

<!-- lance-gatekeeper-fix:v1 agent=7893f7a181fd8bc1ad00acc62d1a85c2
generation=1 -->

---------

Co-authored-by: Gatefixer <313497061+lancedb-gatefixer[bot]@users.noreply.github.com>
Co-authored-by: Xuanwo <github@xuanwo.io>

.bumpversion.toml

@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "0.38.0-beta.10"
+current_version = "0.38.0-beta.11"
 parse = """(?x)
     (?P<major>0|[1-9]\\d*)\\.
     (?P<minor>0|[1-9]\\d*)\\.

Cargo.lock

@@ -5402,7 +5402,7 @@ dependencies = [
 
 [[package]]
 name = "lancedb"
-version = "0.38.0-beta.10"
+version = "0.38.0-beta.11"
 dependencies = [
  "ahash",
  "anyhow",
@@ -5490,7 +5490,7 @@ dependencies = [
 
 [[package]]
 name = "lancedb-nodejs"
-version = "0.38.0-beta.10"
+version = "0.38.0-beta.11"
 dependencies = [
  "arrow-array",
  "arrow-buffer",
@@ -5515,7 +5515,7 @@ dependencies = [
 
 [[package]]
 name = "lancedb-python"
-version = "0.38.0-beta.10"
+version = "0.38.0-beta.11"
 dependencies = [
  "arrow",
  "async-trait",

docs/src/java/java.md

@@ -14,7 +14,7 @@ Add the following dependency to your `pom.xml`:
 <dependency>
     <groupId>com.lancedb</groupId>
     <artifactId>lancedb-core</artifactId>
-    <version>0.38.0-beta.10</version>
+    <version>0.38.0-beta.11</version>
 </dependency>
 ```
 

docs/src/js/classes/Table.md

@@ -1292,6 +1292,18 @@ abstract updateFieldMetadata(updates): Promise<UpdateFieldMetadataResult>
 
 Update per-field (column) metadata.
 
+The following keys are treated specially, by convention, and should be
+used when appropriate:
+
+- `lancedb:description`: for a human-readable description of a field.
+- `lancedb:tag:<name>`: for a user-defined key-value tag, where the suffix
+  names the tag category; e.g. `lancedb:tag:model: "clip"`.
+- `lancedb:logical-column`: for a column grouping; e.g. `feature_v1` and
+  `feature_v2` might be in the same logical column.
+- `lancedb:status`: for status options (`production`, `candidate`,
+  `deprecated`, `archived`) to designate the current life cycle state of
+  this column.
+
 #### Parameters
 
 * **updates**: [`FieldMetadataUpdate`](../interfaces/FieldMetadataUpdate.md)[]

docs/src/js/interfaces/FieldMetadataUpdate.md

@@ -17,7 +17,8 @@ metadata: Record<string, null | string>;
 ```
 
 Metadata key/value pairs. Merged into the field's existing metadata by
-default; a value of `null` deletes that key.
+default; a value of `null` deletes that key. See
+[Table.updateFieldMetadata](../classes/Table.md#updatefieldmetadata) for the conventional `lancedb:*` keys.
 
 ***
 

java/lancedb-core/pom.xml

@@ -8,7 +8,7 @@
     <parent>
       <groupId>com.lancedb</groupId>
       <artifactId>lancedb-parent</artifactId>
-      <version>0.38.0-beta.10</version>
+      <version>0.38.0-beta.11</version>
       <relativePath>../pom.xml</relativePath>
     </parent>
 

java/pom.xml

@@ -6,7 +6,7 @@
 
     <groupId>com.lancedb</groupId>
     <artifactId>lancedb-parent</artifactId>
-    <version>0.38.0-beta.10</version>
+    <version>0.38.0-beta.11</version>
     <packaging>pom</packaging>
     <name>${project.artifactId}</name>
     <description>LanceDB Java SDK Parent POM</description>

nodejs/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
 name = "lancedb-nodejs"
 edition.workspace = true
-version = "0.38.0-beta.10"
+version = "0.38.0-beta.11"
 publish = false
 license.workspace = true
 description.workspace = true

nodejs/__test__/arrow.test.ts

@@ -6,6 +6,9 @@ import * as arrow17 from "apache-arrow-17";
 import * as arrow18 from "apache-arrow-18";
 
 import {
+  Field as CurrentField,
+  LargeBinary as CurrentLargeBinary,
+  Schema as CurrentSchema,
   Vector as CurrentVector,
   convertToTable,
   tableFromIPC as currentTableFromIPC,
@@ -36,6 +39,24 @@ function sampleRecords(): Array<Record<string, any>> {
     },
   ];
 }
+
+it("preserves field metadata from a provided schema", async function () {
+  const jsonMetadata = new Map([["ARROW:extension:name", "lance.json"]]);
+  const schema = new CurrentSchema([
+    new CurrentField("meta", new CurrentLargeBinary(), true, jsonMetadata),
+  ]);
+
+  const table = makeArrowTable(
+    [{ meta: Buffer.from(JSON.stringify({ source: "test" })) }],
+    { schema },
+  );
+
+  expect(table.schema.fields[0].metadata).toEqual(jsonMetadata);
+
+  const roundTripped = currentTableFromIPC(await fromTableToBuffer(table));
+  expect(roundTripped.schema.fields[0].metadata).toEqual(jsonMetadata);
+});
+
 describe.each([arrow15, arrow16, arrow17, arrow18])(
   "Arrow",
   (

nodejs/__test__/embedding.test.ts

@@ -187,6 +187,58 @@ describe("embedding functions", () => {
     const vector0 = JSON.parse(JSON.stringify(arr[0].vector));
     expect(vector0).toEqual([1, 2, 3]);
   });
+  it("should append multiple Python embeddings with the same alias", async () => {
+    @register("python-mock")
+    // biome-ignore lint/correctness/noUnusedVariables: the decorator registers this class
+    class MockEmbeddingFunction extends EmbeddingFunction<string> {
+      ndims() {
+        return 3;
+      }
+      embeddingDataType(): Float {
+        return new Float32();
+      }
+      async computeQueryEmbeddings(_data: string) {
+        return [1, 2, 3];
+      }
+      async computeSourceEmbeddings(data: string[]) {
+        return data.map((value) =>
+          value === "hello world" ? [1, 2, 3] : [4, 5, 6],
+        );
+      }
+    }
+
+    const metadata = new Map([
+      [
+        "embedding_functions",
+        '[{"source_column":"text1","vector_column":"vector1","name":"python-mock","model":{}},{"source_column":"text2","vector_column":"vector2","name":"python-mock","model":{}}]',
+      ],
+    ]);
+    const schema = new Schema(
+      [
+        new Field("text1", new Utf8(), true),
+        new Field("text2", new Utf8(), true),
+        new Field(
+          "vector1",
+          new FixedSizeList(3, new Field("item", new Float32(), true)),
+          true,
+        ),
+        new Field(
+          "vector2",
+          new FixedSizeList(3, new Field("item", new Float32(), true)),
+          true,
+        ),
+      ],
+      metadata,
+    );
+
+    const db = await connect(tmpDir.name);
+    const table = await db.createEmptyTable("test", schema);
+    await table.add([{ text1: "hello world", text2: "goodbye world" }]);
+
+    const rows = await table.query().toArray();
+    expect(JSON.parse(JSON.stringify(rows[0].vector1))).toEqual([1, 2, 3]);
+    expect(JSON.parse(JSON.stringify(rows[0].vector2))).toEqual([4, 5, 6]);
+  });
 
   it("should append generated vectors to a non-nullable schema", async () => {
     @register("non_nullable_schema_test")

nodejs/__test__/table.test.ts

@@ -3561,6 +3561,27 @@ describe("when creating an empty table", () => {
     expect((actualSchema.fields[1].type as Float64).precision).toBe(2);
   });
 
+  it("can add and query JSON data", async () => {
+    const schema = new Schema([
+      new Field("id", new Int32(), true),
+      new Field(
+        "meta",
+        new Utf8(),
+        true,
+        new Map([["ARROW:extension:name", "arrow.json"]]),
+      ),
+    ]);
+    const table = await con.createEmptyTable("json", schema);
+    const meta = JSON.stringify({ x: 1 });
+
+    await table.add([{ id: 1, meta }]);
+
+    const rows = await table.query().toArray();
+    expect(rows).toHaveLength(1);
+    expect(rows[0].id).toBe(1);
+    expect(rows[0].meta).toBe(meta);
+  });
+
   it("can create an empty table from schema that specifies field types by name", async () => {
     const schemaLike = {
       fields: [

nodejs/examples/basic.test.ts

@@ -170,7 +170,7 @@ test("basic table examples", async () => {
     // --8<-- [end:create_index]
 
     // --8<-- [start:delete_rows]
-    await tbl.delete('item = "fizz"');
+    await tbl.delete("item = 'fizz'");
     // --8<-- [end:delete_rows]
 
     // --8<-- [start:drop_table]

nodejs/lancedb/query.ts

@@ -727,11 +727,11 @@ export class VectorQuery extends StandardQueryBase<NativeVectorQuery> {
    * Add a query vector to the search
    *
    * This method can be called multiple times to add multiple query vectors
-   * to the search. If multiple query vectors are added, then they will be searched
-   * in parallel, and the results will be concatenated. A column called `query_index`
-   * will be added to indicate the index of the query vector that produced the result.
-   *
-   * Performance wise, this is equivalent to running multiple queries concurrently.
+   * to the search. A column called `query_index` will be added to indicate the index
+   * of the query vector that produced the result. Flat searches share one table scan
+   * across the query vectors, avoiding the scan and memory amplification of running
+   * multiple queries concurrently. Indexed searches may still perform per-vector
+   * index work.
    */
   addQueryVector(vector: IntoVector): VectorQuery {
     if (vector instanceof Promise) {

nodejs/lancedb/schema.ts

@@ -406,10 +406,11 @@ function matchingFields(fields: Field[], tree: FieldTree): Field[] {
           field.name,
           new Struct(matchingFields(struct.children, value)),
           field.nullable,
+          field.metadata,
         ),
       );
     } else {
-      matches.push(new Field(field.name, value as DataType, field.nullable));
+      matches.push(field);
     }
   }
   return matches;

nodejs/lancedb/table.ts

@@ -630,6 +630,18 @@ export abstract class Table {
 
   /**
    * Update per-field (column) metadata.
+   *
+   * The following keys are treated specially, by convention, and should be
+   * used when appropriate:
+   *
+   * - `lancedb:description`: for a human-readable description of a field.
+   * - `lancedb:tag:<name>`: for a user-defined key-value tag, where the suffix
+   *   names the tag category; e.g. `lancedb:tag:model: "clip"`.
+   * - `lancedb:logical-column`: for a column grouping; e.g. `feature_v1` and
+   *   `feature_v2` might be in the same logical column.
+   * - `lancedb:status`: for status options (`production`, `candidate`,
+   *   `deprecated`, `archived`) to designate the current life cycle state of
+   *   this column.
    * @param {FieldMetadataUpdate[]} updates One or more per-field updates. Each
    * update's metadata is merged into the field's existing metadata by default;
    * a value of `null` deletes that key, and `replace: true` swaps the whole map.
@@ -1555,7 +1567,8 @@ export interface FieldMetadataUpdate {
   path: string;
   /**
    * Metadata key/value pairs. Merged into the field's existing metadata by
-   * default; a value of `null` deletes that key.
+   * default; a value of `null` deletes that key. See
+   * {@link Table.updateFieldMetadata} for the conventional `lancedb:*` keys.
    */
   metadata: Record<string, string | null>;
   /** If true, replace the field's entire metadata map instead of merging. */

nodejs/npm/darwin-arm64/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-darwin-arm64",
-	"version": "0.38.0-beta.10",
+	"version": "0.38.0-beta.11",
 	"os": ["darwin"],
 	"cpu": ["arm64"],
 	"main": "lancedb.darwin-arm64.node",

nodejs/npm/linux-arm64-gnu/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-arm64-gnu",
-	"version": "0.38.0-beta.10",
+	"version": "0.38.0-beta.11",
 	"os": ["linux"],
 	"cpu": ["arm64"],
 	"main": "lancedb.linux-arm64-gnu.node",

nodejs/npm/linux-arm64-musl/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-arm64-musl",
-	"version": "0.38.0-beta.10",
+	"version": "0.38.0-beta.11",
 	"os": ["linux"],
 	"cpu": ["arm64"],
 	"main": "lancedb.linux-arm64-musl.node",

nodejs/npm/linux-x64-gnu/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-x64-gnu",
-	"version": "0.38.0-beta.10",
+	"version": "0.38.0-beta.11",
 	"os": ["linux"],
 	"cpu": ["x64"],
 	"main": "lancedb.linux-x64-gnu.node",

nodejs/npm/linux-x64-musl/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-x64-musl",
-	"version": "0.38.0-beta.10",
+	"version": "0.38.0-beta.11",
 	"os": ["linux"],
 	"cpu": ["x64"],
 	"main": "lancedb.linux-x64-musl.node",

nodejs/npm/win32-arm64-msvc/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lancedb/lancedb-win32-arm64-msvc",
-  "version": "0.38.0-beta.10",
+  "version": "0.38.0-beta.11",
   "os": [
     "win32"
   ],

nodejs/npm/win32-x64-msvc/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-win32-x64-msvc",
-	"version": "0.38.0-beta.10",
+	"version": "0.38.0-beta.11",
 	"os": ["win32"],
 	"cpu": ["x64"],
 	"main": "lancedb.win32-x64-msvc.node",

nodejs/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "@lancedb/lancedb",
-  "version": "0.38.0-beta.10",
+  "version": "0.38.0-beta.11",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@lancedb/lancedb",
-      "version": "0.38.0-beta.10",
+      "version": "0.38.0-beta.11",
       "cpu": [
         "x64",
         "arm64"

nodejs/package.json

@@ -11,7 +11,7 @@
     "ann"
   ],
   "private": false,
-  "version": "0.38.0-beta.10",
+  "version": "0.38.0-beta.11",
   "main": "dist/index.js",
   "exports": {
     ".": "./dist/index.js",

python/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "lancedb-python"
-version = "0.38.0-beta.10"
+version = "0.38.0-beta.11"
 publish = false
 edition.workspace = true
 description = "Python bindings for LanceDB"

python/python/lancedb/query.py

@@ -3401,9 +3401,10 @@ class AsyncQuery(AsyncStandardQuery):
         pass in multiple vectors. When multiple vectors are passed in, if the vector
         column is with multivector type, then the vectors will be treated as a single
         query. Or the vectors will be treated as multiple queries, this can be useful
-        if you want to find the nearest vectors to multiple query vectors.
-        This is not expected to be faster than making multiple queries concurrently;
-        it is just a convenience method. If multiple vectors are passed in then
+        if you want to find the nearest vectors to multiple query vectors. Flat
+        searches share one table scan across the query vectors, avoiding the scan
+        and memory amplification of making multiple queries concurrently. If
+        multiple vectors are passed in then
         an additional column `query_index` will be added to the results. This column
         will contain the index of the query vector that the result is nearest to.
         """
@@ -3532,8 +3533,8 @@ class AsyncFTSQuery(AsyncStandardQuery):
 
         Typically, a single vector is passed in as the query. However, you can also
         pass in multiple vectors.  This can be useful if you want to find the nearest
-        vectors to multiple query vectors. This is not expected to be faster than
-        making multiple queries concurrently; it is just a convenience method.
+        vectors to multiple query vectors. Flat searches share one table scan across
+        the query vectors instead of issuing concurrent full scans.
         If multiple vectors are passed in then an additional column `query_index`
         will be added to the results.  This column will contain the index of the
         query vector that the result is nearest to.

python/python/lancedb/table.py

@@ -2127,12 +2127,25 @@ class Table(ABC):
         ----------
         updates : dict
             One or more dicts, each with:
+
             - "path": str — dot-path to the field (e.g. "embedding" or "a.b.c").
             - "metadata": dict[str, str | None] — keys to set; a value of ``None``
               deletes that key.
             - "replace": bool, optional — replace the field's whole metadata map
               instead of merging (default False).
 
+            The following keys are treated specially, by convention, and should
+            be used when appropriate:
+
+            - "lancedb:description": for a human-readable description of a field.
+            - ``"lancedb:tag:<name>"`` for a user-defined key-value tag, where the
+                suffix names the tag category; e.g. "lancedb:tag:model": "clip".
+            - "lancedb:logical-column" for a column grouping; e.g. "feature_v1"
+                and "feature_v2" might be in the same logical column.
+            - "lancedb:status" for status options ("production", "candidate",
+                "deprecated", "archived") to designate the current life cycle
+                state of this column.
+
         Returns
         -------
         UpdateFieldMetadataResult

python/python/tests/docs/test_basic.py

@@ -105,7 +105,7 @@ def test_quickstart(tmp_path):
     tbl.create_index(num_sub_vectors=1)
     # --8<-- [end:create_index]
     # --8<-- [start:delete_rows]
-    tbl.delete('item = "fizz"')
+    tbl.delete("item = 'fizz'")
     # --8<-- [end:delete_rows]
     # --8<-- [start:drop_table]
     db.drop_table("my_table")
@@ -201,7 +201,7 @@ async def test_quickstart_async(tmp_path):
     await tbl.create_index("vector")
     # --8<-- [end:create_index_async]
     # --8<-- [start:delete_rows_async]
-    await tbl.delete('item = "fizz"')
+    await tbl.delete("item = 'fizz'")
     # --8<-- [end:delete_rows_async]
     # --8<-- [start:drop_table_async]
     await db.drop_table("my_table_async")

python/python/tests/docs/test_guide_tables.py

@@ -266,7 +266,7 @@ def test_table():
     tbl.add(pydantic_model_items)
     # --8<-- [end:add_table_from_pydantic]
     # --8<-- [start:delete_row]
-    tbl.delete('item = "fizz"')
+    tbl.delete("item = 'fizz'")
     # --8<-- [end:delete_row]
     # --8<-- [start:delete_specific_row]
     data = [
@@ -538,7 +538,7 @@ async def test_table_async():
     await async_tbl.add(pydantic_model_items)
     # --8<-- [end:add_table_async_from_pydantic]
     # --8<-- [start:delete_row_async]
-    await async_tbl.delete('item = "fizz"')
+    await async_tbl.delete("item = 'fizz'")
     # --8<-- [end:delete_row_async]
     # --8<-- [start:delete_specific_row_async]
     data = [

python/python/tests/test_query.py

@@ -897,6 +897,23 @@ def test_query_builder_batches(table):
     assert rs_list["id"][1] == 2
 
 
+def test_batch_vector_query_shares_filtered_flat_scan(table):
+    query = (
+        table.search([[1.0, 2.0], [3.0, 4.0]])
+        .where("id > 0", prefilter=True)
+        .limit(1)
+        .select(["id"])
+    )
+
+    plan = query.explain_plan(verbose=True)
+    assert "KNNVectorDistance: queries=2" in plan
+    assert "UnionExec" not in plan
+
+    results = query.to_arrow()
+    assert len(results) == 2
+    assert results["query_index"].to_pylist() == [0, 1]
+
+
 def test_dynamic_projection(table):
     rs = (
         LanceVectorQueryBuilder(table, [0, 0], "vector")

rust/lancedb/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "lancedb"
-version = "0.38.0-beta.10"
+version = "0.38.0-beta.11"
 edition.workspace = true
 description = "LanceDB: A serverless, low-latency vector database for AI applications"
 license.workspace = true

rust/lancedb/src/expr.rs

@@ -19,6 +19,7 @@
 
 mod sql;
 
+pub(crate) use sql::canonicalize_sql_predicate;
 pub use sql::expr_to_sql_string;
 
 use std::sync::Arc;

rust/lancedb/src/expr/sql.rs

@@ -1,10 +1,16 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
+use std::any::TypeId;
+
 use datafusion_common::ScalarValue;
 use datafusion_common::tree_node::{Transformed, TreeNode, TreeNodeRecursion};
 use datafusion_expr::Expr;
-use datafusion_sql::unparser::{self, dialect::Dialect};
+use datafusion_sql::sqlparser::{
+    dialect::{Dialect as SqlParserDialect, GenericDialect},
+    tokenizer::{Token, Tokenizer},
+};
+use datafusion_sql::unparser::{self, dialect::Dialect as UnparserDialect};
 
 /// Unparser dialect that matches the quoting style expected by the Lance SQL
 /// parser.  Lance uses backtick (`` ` ``) as the only delimited-identifier
@@ -19,7 +25,7 @@ use datafusion_sql::unparser::{self, dialect::Dialect};
 ///   lower-case by the SQL parser, which would break case-sensitive schemas).
 struct LanceSqlDialect;
 
-impl Dialect for LanceSqlDialect {
+impl UnparserDialect for LanceSqlDialect {
     fn identifier_quote_style(&self, identifier: &str) -> Option<char> {
         let needs_quote = identifier.chars().any(|c| c.is_ascii_uppercase())
             || !identifier
@@ -30,6 +36,61 @@ impl Dialect for LanceSqlDialect {
     }
 }
 
+/// Lance's tokenizer dialect with SQL-standard double-quoted identifiers added.
+///
+/// Keep this deliberately small: Lance's parser wraps `GenericDialect` and
+/// delegates only identifier recognition, leaving every other dialect option at
+/// its default. In particular, `/*! ... */` remains an ordinary block comment.
+#[derive(Debug, Default)]
+struct PredicateDialect(GenericDialect);
+
+impl SqlParserDialect for PredicateDialect {
+    fn dialect(&self) -> TypeId {
+        self.0.dialect()
+    }
+
+    fn is_identifier_start(&self, ch: char) -> bool {
+        self.0.is_identifier_start(ch)
+    }
+
+    fn is_identifier_part(&self, ch: char) -> bool {
+        self.0.is_identifier_part(ch)
+    }
+
+    fn is_delimited_identifier_start(&self, ch: char) -> bool {
+        ch == '"' || ch == '`'
+    }
+}
+
+/// Canonicalize a raw SQL predicate for Lance's parser.
+///
+/// Lance wraps [`GenericDialect`] for identifier recognition while retaining the
+/// default dialect behavior for every other lexical option. [`PredicateDialect`]
+/// mirrors that contract and additionally recognizes `"` as an identifier
+/// delimiter, allowing this function to rewrite only those identifier tokens.
+pub fn canonicalize_sql_predicate(predicate: &str) -> crate::Result<String> {
+    let dialect = PredicateDialect::default();
+    let tokens = Tokenizer::new(&dialect, predicate)
+        .with_unescape(false)
+        .tokenize()
+        .map_err(|err| crate::Error::InvalidInput {
+            message: format!("invalid SQL predicate: {err}"),
+        })?;
+
+    Ok(tokens
+        .into_iter()
+        .map(|token| match token {
+            Token::Word(word) if word.quote_style == Some('"') => {
+                // with_unescape(false) retains doubled double quotes. Decode
+                // those before escaping any backticks for Lance's delimiter.
+                let identifier = word.value.replace("\"\"", "\"").replace('`', "``");
+                format!("`{identifier}`")
+            }
+            other => other.to_string(),
+        })
+        .collect())
+}
+
 /// Prefix for placeholder strings inserted in place of binary literals.  Chosen
 /// to be extremely unlikely to occur in user data.
 const BINARY_PLACEHOLDER_PREFIX: &str = "__lancedb_binary_placeholder_";
@@ -113,3 +174,51 @@ pub fn expr_to_sql_string(expr: &Expr) -> crate::Result<String> {
     }
     Ok(sql)
 }
+
+#[cfg(test)]
+mod tests {
+    use super::canonicalize_sql_predicate;
+
+    #[test]
+    fn normalizes_double_quoted_identifiers() {
+        assert_eq!(
+            canonicalize_sql_predicate(r#""PartyAbbrev" = 'D'"#).unwrap(),
+            "`PartyAbbrev` = 'D'"
+        );
+        assert_eq!(
+            canonicalize_sql_predicate(r#""MetaData"."userId" = 5"#).unwrap(),
+            "`MetaData`.`userId` = 5"
+        );
+        assert_eq!(
+            canonicalize_sql_predicate(r#""a""b" = 1"#).unwrap(),
+            "`a\"b` = 1"
+        );
+    }
+
+    #[test]
+    fn preserves_quotes_inside_literals_and_backticks() {
+        let filter = r#"name = 'Alice "Ace"' AND `quoted"field` = 1"#;
+        assert_eq!(canonicalize_sql_predicate(filter).unwrap(), filter);
+    }
+
+    #[test]
+    fn preserves_literals_and_comments_using_lance_dialect_rules() {
+        let predicate = r#"path = '\' AND "PartyAbbrev" = 'D' -- unmatched " in comment"#;
+        assert_eq!(
+            canonicalize_sql_predicate(predicate).unwrap(),
+            r#"path = '\' AND `PartyAbbrev` = 'D' -- unmatched " in comment"#
+        );
+
+        let predicate = r#"id = 1 /* unmatched " in block comment */"#;
+        assert_eq!(canonicalize_sql_predicate(predicate).unwrap(), predicate);
+
+        let predicate = r#"id = 1 /*! OR "PartyAbbrev" = 'D' */"#;
+        assert_eq!(canonicalize_sql_predicate(predicate).unwrap(), predicate);
+    }
+
+    #[test]
+    fn rejects_unterminated_double_quoted_identifier() {
+        let error = canonicalize_sql_predicate(r#""PartyAbbrev = 'D'"#).unwrap_err();
+        assert!(matches!(error, crate::Error::InvalidInput { .. }));
+    }
+}

rust/lancedb/src/materialized_view.rs

@@ -170,6 +170,15 @@ pub(crate) fn plan(
     filter: Option<&str>,
     limit: Option<u64>,
 ) -> Result<(MaterializedViewDefinition, Vec<ArrowField>, Lineage)> {
+    let filter = filter
+        .map(crate::expr::canonicalize_sql_predicate)
+        .transpose()
+        .map_err(|err| match err {
+            Error::InvalidInput { message } => Error::InvalidInput {
+                message: format!("invalid view filter: {message}"),
+            },
+            err => err,
+        })?;
     let projections: Vec<(String, String)> = if projections.is_empty() {
         source_schema
             .fields()
@@ -274,7 +283,7 @@ pub(crate) fn plan(
         declared.push(output);
     }
 
-    if let Some(filter) = filter {
+    if let Some(filter) = filter.as_deref() {
         let expr = planner
             .parse_filter(filter)
             .map_err(|e| Error::InvalidInput {
@@ -314,7 +323,7 @@ pub(crate) fn plan(
             .into_iter()
             .map(|(output, expression)| ViewProjection { output, expression })
             .collect(),
-        filter: filter.map(String::from),
+        filter,
         limit,
         inputs,
     };

rust/lancedb/src/materialized_view/refresh.rs

@@ -46,8 +46,9 @@ use lance_table::format::Fragment;
 use serde::{Deserialize, Serialize};
 
 use super::{
-    INCARNATION_META_KEY, MaterializedViewDefinition, REFRESHED_AT_MS_META_KEY,
-    SOURCE_ROW_ID_COLUMN, SOURCE_VERSION_META_KEY,
+    DEFINITION_META_KEY, INCARNATION_META_KEY, MaterializedViewDefinition,
+    REFRESHED_AT_MS_META_KEY, SOURCE_ROW_ID_COLUMN, SOURCE_VERSION_META_KEY,
+    definition_to_metadata,
 };
 use crate::database::OpenTableRequest;
 use crate::table::{NativeTable, NativeTableExt, Table};
@@ -197,8 +198,28 @@ pub(crate) async fn execute_refresh(
             ),
         });
     }
+    let definition_changed =
+        definition.filter != replanned.filter || definition.inputs != replanned.inputs;
     let definition = &replanned;
 
+    // A watermark written for a legacy raw filter certifies the rows that
+    // filter produced, not the canonical predicate above. Rebuild instead of
+    // accepting or advancing it, and persist the migrated definition in the
+    // same metadata commit that certifies the replacement rows.
+    if definition_changed {
+        return rebuild(
+            view_native,
+            &view_ds,
+            &source_ds,
+            source_version,
+            source_ts,
+            definition,
+            true,
+            expected_incarnation,
+        )
+        .await;
+    }
+
     let metadata = &view_ds.schema().metadata;
     let watermark: Option<u64> = metadata
         .get(SOURCE_VERSION_META_KEY)
@@ -257,6 +278,7 @@ pub(crate) async fn execute_refresh(
                         source_version,
                         source_ts,
                         definition,
+                        false,
                         expected_incarnation,
                     )
                     .await
@@ -271,6 +293,7 @@ pub(crate) async fn execute_refresh(
                 source_version,
                 source_ts,
                 definition,
+                false,
                 expected_incarnation,
             )
             .await
@@ -683,6 +706,7 @@ async fn incremental(
             view_ds.clone(),
             source_version,
             source_ts,
+            None,
             expected_incarnation,
         )
         .await?;
@@ -704,6 +728,7 @@ async fn incremental(
             published,
             source_version,
             source_ts,
+            None,
             expected_incarnation,
         )
         .await?;
@@ -775,6 +800,7 @@ async fn incremental(
             published,
             source_version,
             source_ts,
+            None,
             expected_incarnation,
         )
         .await?;
@@ -824,12 +850,14 @@ async fn incremental(
         appended,
         source_version,
         source_ts,
+        None,
         expected_incarnation,
     )
     .await?;
     Ok(Some(result))
 }
 
+#[allow(clippy::too_many_arguments)]
 async fn rebuild(
     view_native: &NativeTable,
     view_ds: &Dataset,
@@ -837,6 +865,7 @@ async fn rebuild(
     source_version: u64,
     source_ts: u128,
     definition: &MaterializedViewDefinition,
+    persist_definition: bool,
     expected_incarnation: Option<&str>,
 ) -> Result<RefreshMaterializedViewResult> {
     let rows_written = Arc::new(AtomicU64::new(0));
@@ -867,6 +896,7 @@ async fn rebuild(
         replaced,
         source_version,
         source_ts,
+        persist_definition.then_some(definition),
         expected_incarnation,
     )
     .await?;
@@ -981,6 +1011,7 @@ async fn stamp_watermark(
     mut dataset: Dataset,
     source_version: u64,
     source_ts: u128,
+    definition: Option<&MaterializedViewDefinition>,
     expected_incarnation: Option<&str>,
 ) -> Result<u64> {
     ensure_incarnation(&dataset, expected_incarnation, dataset.uri()).await?;
@@ -993,27 +1024,32 @@ async fn stamp_watermark(
         .get(INCARNATION_META_KEY)
         .cloned()
         .unwrap_or_else(|| uuid::Uuid::new_v4().to_string());
-    dataset
-        .update_schema_metadata([
-            (INCARNATION_META_KEY.to_string(), Some(incarnation)),
-            (
-                SOURCE_VERSION_META_KEY.to_string(),
-                Some(source_version.to_string()),
-            ),
-            (
-                SOURCE_VERSION_TS_META_KEY.to_string(),
-                Some(source_ts.to_string()),
-            ),
-            (
-                REFRESHED_AT_MS_META_KEY.to_string(),
-                Some(now_ms().to_string()),
-            ),
-            (
-                VIEW_VERSION_META_KEY.to_string(),
-                Some(predicted.to_string()),
-            ),
-        ])
-        .await?;
+    let mut metadata = vec![(INCARNATION_META_KEY.to_string(), Some(incarnation))];
+    if let Some(definition) = definition {
+        metadata.push((
+            DEFINITION_META_KEY.to_string(),
+            Some(definition_to_metadata(definition)?),
+        ));
+    }
+    metadata.extend([
+        (
+            SOURCE_VERSION_META_KEY.to_string(),
+            Some(source_version.to_string()),
+        ),
+        (
+            SOURCE_VERSION_TS_META_KEY.to_string(),
+            Some(source_ts.to_string()),
+        ),
+        (
+            REFRESHED_AT_MS_META_KEY.to_string(),
+            Some(now_ms().to_string()),
+        ),
+        (
+            VIEW_VERSION_META_KEY.to_string(),
+            Some(predicted.to_string()),
+        ),
+    ]);
+    dataset.update_schema_metadata(metadata).await?;
     let actual = dataset.version().version;
     if actual != predicted {
         return Err(Error::Runtime {
@@ -1585,6 +1621,106 @@ mod tests {
         assert_eq!(read(view.table(), "x").await, vec![20, 40]);
     }
 
+    #[tokio::test]
+    async fn test_mixed_case_filter_is_canonicalized_for_lineage_and_refresh() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let batch = record_batch!(
+            ("id", Int32, [1, 2, 3]),
+            ("PartyAbbrev", Utf8, ["D", "R", "D"])
+        )
+        .unwrap();
+        conn.create_table("src", batch)
+            .write_options(crate::materialized_view::tests::stable_row_ids())
+            .execute()
+            .await
+            .unwrap();
+        conn.create_materialized_view("democrats", "src")
+            .select([("id", "id")])
+            .only_if(r#""PartyAbbrev" = 'D'"#)
+            .execute()
+            .await
+            .unwrap();
+
+        // Reopen from schema metadata so these assertions cover the stored
+        // predicate and lineage, not only the declaration-time handle.
+        let view = conn.open_materialized_view("democrats").await.unwrap();
+        assert_eq!(
+            view.definition().filter.as_deref(),
+            Some("`PartyAbbrev` = 'D'")
+        );
+        assert_eq!(view.definition().inputs, ["PartyAbbrev", "id"]);
+
+        let result = view.refresh().execute().await.unwrap();
+        assert_eq!(result.rows_written, 2);
+        assert_eq!(read(view.table(), "id").await, vec![1, 3]);
+    }
+
+    #[tokio::test]
+    async fn test_legacy_raw_filter_rebuilds_and_persists_canonical_definition() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let batch = record_batch!(
+            ("id", Int32, [1, 2, 3]),
+            ("PartyAbbrev", Utf8, ["D", "R", "D"])
+        )
+        .unwrap();
+        conn.create_table("legacy_src", batch)
+            .write_options(crate::materialized_view::tests::stable_row_ids())
+            .execute()
+            .await
+            .unwrap();
+        let view = conn
+            .create_materialized_view("legacy_view", "legacy_src")
+            .select([("id", "id")])
+            .only_if(r#""PartyAbbrev" = 'X'"#)
+            .execute()
+            .await
+            .unwrap();
+        assert_eq!(view.refresh().execute().await.unwrap().rows_written, 0);
+
+        // Model a definition and up-to-date watermark written before filter
+        // canonicalization was applied to materialized views.
+        let mut legacy = view.definition().clone();
+        legacy.filter = Some(r#""PartyAbbrev" = 'D'"#.into());
+        legacy.inputs = vec!["id".into()];
+        let native = view.table().as_native().unwrap();
+        let mut dataset = native.dataset.get().await.unwrap().as_ref().clone();
+        let predicted = dataset.version().version + 1;
+        dataset
+            .update_schema_metadata([
+                (
+                    DEFINITION_META_KEY.to_string(),
+                    Some(definition_to_metadata(&legacy).unwrap()),
+                ),
+                (
+                    VIEW_VERSION_META_KEY.to_string(),
+                    Some(predicted.to_string()),
+                ),
+            ])
+            .await
+            .unwrap();
+        native.dataset.update(dataset);
+
+        let reopened = conn.open_materialized_view("legacy_view").await.unwrap();
+        let result = reopened.refresh().execute().await.unwrap();
+        assert_eq!(result.mode, RefreshMode::Rebuild);
+        assert_eq!(result.rows_written, 2);
+        assert_eq!(read(reopened.table(), "id").await, vec![1, 3]);
+
+        // A fresh handle proves the migration was stored alongside the new
+        // watermark and therefore happens only once.
+        let migrated = conn.open_materialized_view("legacy_view").await.unwrap();
+        assert_eq!(
+            migrated.definition().filter.as_deref(),
+            Some("`PartyAbbrev` = 'D'")
+        );
+        assert_eq!(migrated.definition().inputs, ["PartyAbbrev", "id"]);
+        assert_eq!(
+            migrated.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+        assert_eq!(read(migrated.table(), "id").await, vec![1, 3]);
+    }
+
     #[tokio::test]
     async fn test_append_refreshes_incrementally() {
         let (_conn, source, view) = refreshed_doubled(vec![1, 2]).await;
@@ -2767,7 +2903,7 @@ mod tests {
         let stale = view_native.dataset.get().await.unwrap().as_ref().clone();
         view.table().delete("x = 1").await.unwrap();
 
-        let err = stamp_watermark(view_native, stale, 99, 99, None).await;
+        let err = stamp_watermark(view_native, stale, 99, 99, None, None).await;
         assert!(err.is_err());
 
         let result = view.refresh().execute().await.unwrap();

rust/lancedb/src/query.rs

@@ -399,6 +399,9 @@ pub trait QueryBase {
     /// x > 5 OR y = 'test'
     /// ```
     ///
+    /// Identifiers may be delimited with SQL-standard double quotes or
+    /// backticks. String literals must use single quotes.
+    ///
     /// Filtering performance can often be improved by creating a scalar index
     /// on the filter column(s).
     ///
@@ -913,6 +916,17 @@ impl QueryRequest {
     /// use different representations) the error is recorded and surfaced later
     /// by [`Self::check_filter`].
     pub(crate) fn add_filter(&mut self, new: QueryFilter) {
+        let new = match new {
+            QueryFilter::Sql(filter) => match crate::expr::canonicalize_sql_predicate(&filter) {
+                Ok(filter) => QueryFilter::Sql(filter),
+                Err(err) => {
+                    self.filter_error = Some(err.to_string());
+                    return;
+                }
+            },
+            other => other,
+        };
+
         self.filter = Some(match self.filter.take() {
             None => new,
             Some(existing) => match and_filters(existing, new) {
@@ -1174,12 +1188,12 @@ impl VectorQuery {
 
     /// Add another query vector to the search.
     ///
-    /// Multiple searches will be dispatched as part of the query.
-    /// This is a convenience method for adding multiple query vectors
-    /// to the search. It is not expected to be faster than issuing
-    /// multiple queries concurrently.
+    /// Multiple searches will be dispatched as a batch. Flat searches share
+    /// one table scan across the query vectors, avoiding the scan and memory
+    /// amplification of issuing the searches concurrently. Indexed searches
+    /// may still perform per-vector index work.
     ///
-    /// The output data will contain an additional columns `query_index` which
+    /// The output data will contain an additional column `query_index` which
     /// will contain the index of the query vector that was used to generate the
     /// result.
     pub fn add_query_vector(mut self, vector: impl IntoQueryVector) -> Result<Self> {
@@ -1646,10 +1660,14 @@ mod tests {
     use std::{collections::HashSet, sync::Arc};
 
     use super::*;
-    use arrow::{array::downcast_array, compute::concat_batches, datatypes::Int32Type};
+    use arrow::{
+        array::downcast_array,
+        compute::concat_batches,
+        datatypes::{Int32Type, UInt8Type},
+    };
     use arrow_array::{
-        FixedSizeListArray, Float32Array, Int32Array, RecordBatch, StringArray, cast::AsArray,
-        types::Float32Type,
+        FixedSizeListArray, Float32Array, Int32Array, RecordBatch, RecordBatchIterator,
+        StringArray, cast::AsArray, types::Float32Type,
     };
     use arrow_schema::{DataType, Field as ArrowField, Schema as ArrowSchema};
     use futures::{StreamExt, TryStreamExt};
@@ -1878,6 +1896,157 @@ mod tests {
         query.execute().await.unwrap();
     }
 
+    #[tokio::test]
+    async fn test_double_quoted_predicates_across_table_operations() {
+        let tmp_dir = tempdir().unwrap();
+        let dataset_path = tmp_dir.path().join("test.lance");
+        let uri = dataset_path.to_str().unwrap();
+        let schema = Arc::new(ArrowSchema::new(vec![
+            ArrowField::new("id", DataType::Int32, false),
+            ArrowField::new("PartyAbbrev", DataType::Utf8, false),
+            ArrowField::new("path", DataType::Utf8, false),
+        ]));
+        let batch = RecordBatch::try_new(
+            schema.clone(),
+            vec![
+                Arc::new(Int32Array::from(vec![1, 2, 3, 4])),
+                Arc::new(StringArray::from(vec!["D", "R", "R", "D"])),
+                Arc::new(StringArray::from(vec!["\\", "\\", "x", "x"])),
+            ],
+        )
+        .unwrap();
+
+        let conn = connect(uri).execute().await.unwrap();
+        let table = conn.create_table("parties", batch).execute().await.unwrap();
+        let batches = table
+            .query()
+            .only_if(r#""PartyAbbrev" = 'D'"#)
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+
+        assert_eq!(batches.iter().map(RecordBatch::num_rows).sum::<usize>(), 2);
+        assert_eq!(
+            table
+                .count_rows(Some(r#""PartyAbbrev" = 'D'"#.to_string()))
+                .await
+                .unwrap(),
+            2
+        );
+
+        // Public BaseTable dispatch cannot bypass canonicalization.
+        let query = AnyQuery::Query(QueryRequest {
+            filter: Some(QueryFilter::Sql(r#""PartyAbbrev" = 'D'"#.to_string())),
+            ..Default::default()
+        });
+        let batches = table
+            .base_table()
+            .query(&query, Default::default())
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        assert_eq!(batches.iter().map(RecordBatch::num_rows).sum::<usize>(), 2);
+        assert_eq!(
+            table
+                .base_table()
+                .count_rows(Some(crate::table::Filter::Sql(
+                    r#""PartyAbbrev" = 'D'"#.to_string(),
+                )))
+                .await
+                .unwrap(),
+            2
+        );
+
+        for predicate in [
+            r#"id = 1 -- unmatched " in a valid SQL comment"#,
+            r#"id = 1 /* unmatched " in a valid SQL comment */"#,
+            r#"id = 1 /*! OR "PartyAbbrev" = 'D' */"#,
+            r#"path = '\' AND "PartyAbbrev" = 'D'"#,
+        ] {
+            let batches = table
+                .query()
+                .only_if(predicate)
+                .execute()
+                .await
+                .unwrap()
+                .try_collect::<Vec<_>>()
+                .await
+                .unwrap();
+            assert_eq!(batches.iter().map(RecordBatch::num_rows).sum::<usize>(), 1);
+        }
+
+        // The same canonical predicate contract applies to both merge filters.
+        let source = RecordBatch::try_new(
+            schema.clone(),
+            vec![
+                Arc::new(Int32Array::from(vec![1, 2, 3])),
+                Arc::new(StringArray::from(vec!["D", "R", "R"])),
+                Arc::new(StringArray::from(vec!["\\", "\\", "x"])),
+            ],
+        )
+        .unwrap();
+        let mut merge = table.merge_insert(&["id"]);
+        merge.when_not_matched_by_source_delete(Some(r#""PartyAbbrev" = 'D'"#.to_string()));
+        let result = table
+            .base_table()
+            .merge_insert(
+                merge,
+                Box::new(RecordBatchIterator::new(vec![Ok(source)], schema.clone())),
+            )
+            .await
+            .unwrap();
+        assert_eq!(result.num_deleted_rows, 1);
+
+        let source = RecordBatch::try_new(
+            schema.clone(),
+            vec![
+                Arc::new(Int32Array::from(vec![1, 2, 3])),
+                Arc::new(StringArray::from(vec!["U", "U", "U"])),
+                Arc::new(StringArray::from(vec!["\\", "\\", "x"])),
+            ],
+        )
+        .unwrap();
+        let mut merge = table.merge_insert(&["id"]);
+        merge.when_matched_update_all(Some(r#"target."PartyAbbrev" = 'D'"#.to_string()));
+        merge
+            .execute(Box::new(RecordBatchIterator::new(vec![Ok(source)], schema)))
+            .await
+            .unwrap();
+        assert_eq!(
+            table
+                .count_rows(Some(r#""PartyAbbrev" = 'U'"#.to_string()))
+                .await
+                .unwrap(),
+            1
+        );
+
+        let update = table
+            .update()
+            .only_if(r#""PartyAbbrev" = 'R'"#)
+            .column("PartyAbbrev", "'X'");
+        table.base_table().update(update).await.unwrap();
+        assert_eq!(
+            table
+                .count_rows(Some(r#""PartyAbbrev" = 'X'"#.to_string()))
+                .await
+                .unwrap(),
+            2
+        );
+
+        let result = table
+            .base_table()
+            .delete(crate::table::Predicate::String(r#""PartyAbbrev" = 'X'"#))
+            .await
+            .unwrap();
+        assert_eq!(result.num_deleted_rows, 2);
+        assert_eq!(table.count_rows(None).await.unwrap(), 1);
+    }
+
     #[tokio::test]
     async fn test_select_with_transform() {
         let batches = make_non_empty_batches();
@@ -2334,7 +2503,8 @@ mod tests {
             .limit(1);
 
         let plan = query.explain_plan(true).await.unwrap();
-        assert!(plan.contains("UnionExec"));
+        assert!(plan.contains("KNNVectorDistance: queries=2"));
+        assert!(!plan.contains("UnionExec"));
 
         let results = query
             .execute()
@@ -2349,6 +2519,100 @@ mod tests {
         // We don't guarantee order.
         assert!(query_index.values().contains(&0));
         assert!(query_index.values().contains(&1));
+
+        // Batch KNN does not support a per-query offset, so offset queries keep
+        // the legacy per-vector plan to preserve their result semantics.
+        let offset_query = table
+            .query()
+            .nearest_to(&[0.1, 0.2, 0.3, 0.4])
+            .unwrap()
+            .add_query_vector(&[0.5, 0.6, 0.7, 0.8])
+            .unwrap()
+            .limit(1)
+            .offset(1);
+        assert!(
+            offset_query
+                .explain_plan(true)
+                .await
+                .unwrap()
+                .contains("UnionExec")
+        );
+        let offset_results = offset_query
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        assert_eq!(
+            offset_results
+                .iter()
+                .map(RecordBatch::num_rows)
+                .sum::<usize>(),
+            2
+        );
+    }
+
+    #[tokio::test]
+    async fn test_multiple_binary_query_vectors() {
+        let vectors = FixedSizeListArray::from_iter_primitive::<UInt8Type, _, _>(
+            vec![
+                Some(vec![Some(0), Some(0)]),
+                Some(vec![Some(255), Some(255)]),
+            ],
+            2,
+        );
+        let schema = Arc::new(ArrowSchema::new(vec![
+            ArrowField::new("id", DataType::Int32, false),
+            ArrowField::new("vector", vectors.data_type().clone(), false),
+        ]));
+        let batch = RecordBatch::try_new(
+            schema,
+            vec![Arc::new(Int32Array::from(vec![0, 1])), Arc::new(vectors)],
+        )
+        .unwrap();
+
+        let conn = connect("memory://").execute().await.unwrap();
+        let table = conn
+            .create_table("binary_batch", batch)
+            .execute()
+            .await
+            .unwrap();
+        let query = table
+            .query()
+            .nearest_to(&[0.0, 0.0])
+            .unwrap()
+            .add_query_vector(&[255.0, 255.0])
+            .unwrap()
+            .distance_type(DistanceType::Hamming)
+            .limit(1);
+
+        // Binary queries retain the per-vector plan because Lance's binary
+        // nearest path requires primitive UInt8 query arrays.
+        assert!(
+            query
+                .explain_plan(true)
+                .await
+                .unwrap()
+                .contains("UnionExec")
+        );
+
+        let results = query
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        let results = concat_batches(&results[0].schema(), &results).unwrap();
+        assert_eq!(results.num_rows(), 2);
+
+        let ids = results["id"].as_primitive::<Int32Type>();
+        assert!(ids.values().contains(&0));
+        assert!(ids.values().contains(&1));
+        let query_index = results["query_index"].as_primitive::<Int32Type>();
+        assert!(query_index.values().contains(&0));
+        assert!(query_index.values().contains(&1));
     }
 
     #[tokio::test]

rust/lancedb/src/remote/table.rs

@@ -1379,10 +1379,11 @@ impl<S: HttpSend> RemoteTable<S> {
         query: &AnyQuery,
         version: Option<u64>,
     ) -> Result<Vec<serde_json::Value>> {
+        let query = query.canonicalized()?;
         let mut base_body = serde_json::json!({ "version": version });
         self.apply_branch_body(&mut base_body);
 
-        match query {
+        match &query {
             AnyQuery::Query(query) => {
                 let mut body = base_body.clone();
                 self.apply_query_params(&mut body, query)?;
@@ -2491,7 +2492,7 @@ impl<S: HttpSend> BaseTable for RemoteTable<S> {
 
         let mut body = if let Some(filter) = filter {
             let filter_sql = match filter {
-                Filter::Sql(sql) => sql.clone(),
+                Filter::Sql(sql) => crate::expr::canonicalize_sql_predicate(&sql)?,
                 Filter::Datafusion(expr) => expr_to_sql_string(&expr)?,
             };
             serde_json::json!({ "predicate": filter_sql, "version": read_snapshot.version })
@@ -2747,7 +2748,8 @@ impl<S: HttpSend> BaseTable for RemoteTable<S> {
         Ok(final_analyze)
     }
 
-    async fn update(&self, update: UpdateBuilder) -> Result<UpdateResult> {
+    async fn update(&self, mut update: UpdateBuilder) -> Result<UpdateResult> {
+        update.canonicalize_filter()?;
         self.check_mutable().await?;
         let request = self
             .client
@@ -2794,7 +2796,7 @@ impl<S: HttpSend> BaseTable for RemoteTable<S> {
     async fn delete(&self, predicate: Predicate<'_>) -> Result<DeleteResult> {
         self.check_mutable().await?;
         let predicate_sql = match predicate {
-            Predicate::String(s) => s.to_string(),
+            Predicate::String(s) => crate::expr::canonicalize_sql_predicate(s)?,
             Predicate::Expr(expr) => expr_to_sql_string(expr)?,
         };
         let mut body = serde_json::json!({ "predicate": predicate_sql });
@@ -2851,9 +2853,10 @@ impl<S: HttpSend> BaseTable for RemoteTable<S> {
 
     async fn merge_insert(
         &self,
-        params: MergeInsertBuilder,
+        mut params: MergeInsertBuilder,
         new_data: Box<dyn RecordBatchReader + Send>,
     ) -> Result<MergeResult> {
+        params.canonicalize_filters()?;
         self.check_mutable().await?;
 
         let timeout = params.timeout;
@@ -3864,13 +3867,17 @@ mod tests {
             );
             assert_eq!(
                 request.body().unwrap().as_bytes().unwrap(),
-                br#"{"predicate":"a > 10","version":null}"#
+                br#"{"predicate":"`A` > 10","version":null}"#
             );
 
             http::Response::builder().status(200).body("42").unwrap()
         });
 
-        let count = table.count_rows(Some("a > 10".into())).await.unwrap();
+        let count = table
+            .base_table()
+            .count_rows(Some(Filter::Sql(r#""A" > 10"#.into())))
+            .await
+            .unwrap();
         assert_eq!(count, 42);
     }
 
@@ -4353,7 +4360,7 @@ mod tests {
                     assert_eq!(expression, "b - 1");
 
                     let only_if = value.get("predicate").unwrap().as_str().unwrap();
-                    assert_eq!(only_if, "b > 10");
+                    assert_eq!(only_if, "`B` > 10");
                 }
 
                 if old_server {
@@ -4369,14 +4376,12 @@ mod tests {
             }
         });
 
-        let result = table
+        let update = table
             .update()
             .column("a", "a + 1")
             .column("b", "b - 1")
-            .only_if("b > 10")
-            .execute()
-            .await
-            .unwrap();
+            .only_if(r#""B" > 10"#);
+        let result = table.base_table().update(update).await.unwrap();
 
         assert_eq!(result.version, if old_server { 0 } else { 43 });
         assert_eq!(result.rows_updated, if old_server { 0 } else { 5 });
@@ -4463,10 +4468,10 @@ mod tests {
 
                 let params = request.url().query_pairs().collect::<HashMap<_, _>>();
                 assert_eq!(params["on"], "some_col");
-                assert_eq!(params["when_matched_update_all"], "false");
+                assert_eq!(params["when_matched_update_all"], "true");
                 assert_eq!(params["when_not_matched_insert_all"], "false");
                 assert_eq!(params["when_not_matched_by_source_delete"], "false");
-                assert!(!params.contains_key("when_matched_update_all_filt"));
+                assert_eq!(params["when_matched_update_all_filt"], "target.`A` > 0");
                 assert!(!params.contains_key("when_not_matched_by_source_delete_filt"));
                 assert!(!params.contains_key("use_index"));
 
@@ -4483,11 +4488,9 @@ mod tests {
             }
         });
 
-        let result = table
-            .merge_insert(&["some_col"])
-            .execute(data)
-            .await
-            .unwrap();
+        let mut merge = table.merge_insert(&["some_col"]);
+        merge.when_matched_update_all(Some(r#"target."A" > 0"#.into()));
+        let result = table.base_table().merge_insert(merge, data).await.unwrap();
 
         assert_eq!(result.version, if old_server { 0 } else { 43 });
         if !old_server {
@@ -4549,7 +4552,7 @@ mod tests {
                 let body = request.body().unwrap().as_bytes().unwrap();
                 let body: serde_json::Value = serde_json::from_slice(body).unwrap();
                 let predicate = body.get("predicate").unwrap().as_str().unwrap();
-                assert_eq!(predicate, "id in (1, 2, 3)");
+                assert_eq!(predicate, "`ID` in (1, 2, 3)");
 
                 if old_server {
                     http::Response::builder()
@@ -4567,7 +4570,11 @@ mod tests {
             }
         });
 
-        let result = table.delete("id in (1, 2, 3)").await.unwrap();
+        let result = table
+            .base_table()
+            .delete(Predicate::String(r#""ID" in (1, 2, 3)"#))
+            .await
+            .unwrap();
         assert_eq!(result.version, if old_server { 0 } else { 43 });
     }
 
@@ -4659,6 +4666,7 @@ mod tests {
             let body = request.body().unwrap().as_bytes().unwrap();
             let body: serde_json::Value = serde_json::from_slice(body).unwrap();
             let expected_body = serde_json::json!({
+                "filter": "`A` > 0",
                 "k": isize::MAX as usize,
                 "prefilter": true,
                 "vector": [], // Empty vector means no vector query.
@@ -4674,9 +4682,13 @@ mod tests {
                 .unwrap()
         });
 
+        let query = AnyQuery::Query(QueryRequest {
+            filter: Some(QueryFilter::Sql(r#""A" > 0"#.into())),
+            ..Default::default()
+        });
         let data = table
-            .query()
-            .execute()
+            .base_table()
+            .query(&query, Default::default())
             .await
             .unwrap()
             .collect::<Vec<_>>()

rust/lancedb/src/table.rs

@@ -1164,7 +1164,10 @@ impl Table {
     ///
     /// * `filter` if present, only count rows matching the filter
     pub async fn count_rows(&self, filter: Option<String>) -> Result<usize> {
-        self.inner.count_rows(filter.map(Filter::Sql)).await
+        let filter = filter
+            .map(|predicate| crate::expr::canonicalize_sql_predicate(&predicate).map(Filter::Sql))
+            .transpose()?;
+        self.inner.count_rows(filter).await
     }
 
     /// Names of the blob v2 columns in this table, in declaration order.
@@ -1364,7 +1367,13 @@ impl Table {
     /// # });
     /// ```
     pub async fn delete(&self, predicate: impl Into<Predicate<'_>>) -> Result<DeleteResult> {
-        self.inner.delete(predicate.into()).await
+        match predicate.into() {
+            Predicate::String(predicate) => {
+                let predicate = crate::expr::canonicalize_sql_predicate(predicate)?;
+                self.inner.delete(Predicate::String(&predicate)).await
+            }
+            predicate @ Predicate::Expr(_) => self.inner.delete(predicate).await,
+        }
     }
 
     /// Create an index on the provided column(s).
@@ -1777,7 +1786,23 @@ impl Table {
         self.inner.alter_columns(alterations).await
     }
 
-    /// Update per-field metadata (merges by default).
+    /// Update per-field (column) metadata.
+    ///
+    /// Each [`FieldMetadataUpdate`] is merged into the field's existing metadata
+    /// by default; use [`FieldMetadataUpdate::remove`] to delete a key, or
+    /// [`FieldMetadataUpdate::replace`] to swap the field's entire metadata map.
+    ///
+    /// The following keys are treated specially, by convention, and should be
+    /// used when appropriate:
+    ///
+    /// - `lancedb:description`: for a human-readable description of a field.
+    /// - `lancedb:tag:<name>`: for a user-defined key-value tag, where the suffix
+    ///   names the tag category; e.g. `lancedb:tag:model: "clip"`.
+    /// - `lancedb:logical-column`: for a column grouping; e.g. `feature_v1` and
+    ///   `feature_v2` might be in the same logical column.
+    /// - `lancedb:status`: for status options (`production`, `candidate`,
+    ///   `deprecated`, `archived`) to designate the current life cycle state of
+    ///   this column.
     pub async fn update_field_metadata(
         &self,
         updates: &[FieldMetadataUpdate],
@@ -3223,7 +3248,10 @@ impl BaseTable for NativeTable {
         let dataset = self.dataset.get().await?;
         match filter {
             None => Ok(dataset.count_rows(None).await?),
-            Some(Filter::Sql(sql)) => Ok(dataset.count_rows(Some(sql)).await?),
+            Some(Filter::Sql(sql)) => {
+                let sql = crate::expr::canonicalize_sql_predicate(&sql)?;
+                Ok(dataset.count_rows(Some(sql)).await?)
+            }
             Some(Filter::Datafusion(_)) => Err(Error::NotSupported {
                 message: "Datafusion filters are not yet supported".to_string(),
             }),

rust/lancedb/src/table/delete.rs

@@ -31,8 +31,9 @@ pub(crate) async fn execute_delete(
     table.dataset.ensure_mutable()?;
     match predicate {
         Predicate::String(s) => {
+            let predicate = crate::expr::canonicalize_sql_predicate(s)?;
             let mut dataset = (*table.dataset.get().await?).clone();
-            let delete_result = dataset.delete(s).boxed().await?;
+            let delete_result = dataset.delete(&predicate).boxed().await?;
             let num_deleted_rows = delete_result.num_deleted_rows;
             let version = dataset.version().version;
             table.dataset.update(dataset);

rust/lancedb/src/table/merge.rs

@@ -220,9 +220,32 @@ impl MergeInsertBuilder {
     ///
     /// Returns version and statistics about the merge operation including the number of rows
     /// inserted, updated, and deleted.
-    pub async fn execute(self, new_data: Box<dyn RecordBatchReader + Send>) -> Result<MergeResult> {
+    pub async fn execute(
+        mut self,
+        new_data: Box<dyn RecordBatchReader + Send>,
+    ) -> Result<MergeResult> {
+        self.canonicalize_filters()?;
         self.table.clone().merge_insert(self, new_data).await
     }
+
+    pub(crate) fn canonicalize_filters(&mut self) -> Result<()> {
+        self.when_matched_update_all_filt =
+            canonicalize_merge_filter(self.when_matched_update_all_filt.take())?;
+        self.when_not_matched_by_source_delete_filt =
+            canonicalize_merge_filter(self.when_not_matched_by_source_delete_filt.take())?;
+        Ok(())
+    }
+}
+
+fn canonicalize_merge_filter(filter: Option<MergeFilter>) -> Result<Option<MergeFilter>> {
+    filter
+        .map(|filter| match filter {
+            MergeFilter::Sql(predicate) => {
+                crate::expr::canonicalize_sql_predicate(&predicate).map(MergeFilter::Sql)
+            }
+            filter @ MergeFilter::Expr(_) => Ok(filter),
+        })
+        .transpose()
 }
 
 /// Internal implementation of the merge insert logic
@@ -230,9 +253,10 @@ impl MergeInsertBuilder {
 /// This logic was moved from NativeTable::merge_insert to keep table.rs clean.
 pub(crate) async fn execute_merge_insert(
     table: &NativeTable,
-    params: MergeInsertBuilder,
+    mut params: MergeInsertBuilder,
     new_data: Box<dyn RecordBatchReader + Send>,
 ) -> Result<MergeResult> {
+    params.canonicalize_filters()?;
     super::computed_columns::ensure_no_function_bindings_for_mutation(
         table.schema().await?.as_ref(),
         "merge_insert",
@@ -1407,4 +1431,195 @@ mod lsm_tests {
             "LSM vector search must rank the memtable row first"
         );
     }
+
+    #[tokio::test]
+    async fn lsm_cosine_distance_scale_and_mixed_tier_ordering() {
+        use arrow::array::{FixedSizeListBuilder, Float32Builder};
+        use arrow::datatypes::Float32Type;
+
+        use crate::index::Index;
+        use crate::index::vector::IvfPqIndexBuilder;
+
+        const DIM: usize = 8;
+        const N: usize = 256;
+
+        fn normalized_vector(state: &mut u64) -> Vec<f32> {
+            let mut vector = (0..DIM)
+                .map(|_| {
+                    *state = state
+                        .wrapping_mul(6_364_136_223_846_793_005)
+                        .wrapping_add(1);
+                    ((*state >> 32) as u32 as f32 / u32::MAX as f32) * 2.0 - 1.0
+                })
+                .collect::<Vec<_>>();
+            let norm = vector.iter().map(|value| value * value).sum::<f32>().sqrt();
+            vector.iter_mut().for_each(|value| *value /= norm);
+            vector
+        }
+
+        let schema = Arc::new(Schema::new(vec![
+            Field::new("id", DataType::Int64, false),
+            Field::new(
+                "vec",
+                DataType::FixedSizeList(
+                    Arc::new(Field::new("item", DataType::Float32, true)),
+                    DIM as i32,
+                ),
+                false,
+            ),
+        ]));
+        let make_batch = |rows: Vec<(i64, Vec<f32>)>| {
+            let ids = rows.iter().map(|(id, _)| *id).collect::<Vec<_>>();
+            let mut vectors = FixedSizeListBuilder::new(Float32Builder::new(), DIM as i32);
+            for (_, vector) in &rows {
+                vectors.values().append_slice(vector);
+                vectors.append(true);
+            }
+            RecordBatch::try_new(
+                schema.clone(),
+                vec![Arc::new(Int64Array::from(ids)), Arc::new(vectors.finish())],
+            )
+            .unwrap()
+        };
+        let first_result = |batches: &[RecordBatch]| {
+            let batch = &batches[0];
+            let id = batch["id"].as_primitive::<Int64Type>().value(0);
+            let distance = batch["_distance"].as_primitive::<Float32Type>().value(0);
+            (id, distance)
+        };
+
+        let mut state = 42;
+        let base_rows = (0..N)
+            .map(|id| (id as i64, normalized_vector(&mut state)))
+            .collect::<Vec<_>>();
+        let query = normalized_vector(&mut state);
+
+        let dir = tempdir().unwrap();
+        let conn = connect(dir.path().to_str().unwrap())
+            .execute()
+            .await
+            .unwrap();
+        let base = make_batch(base_rows);
+        let reader: Box<dyn RecordBatchReader + Send> =
+            Box::new(RecordBatchIterator::new(vec![Ok(base)], schema.clone()));
+        let table = conn
+            .create_table("cosine_lsm", reader)
+            .execute()
+            .await
+            .unwrap();
+        table.set_unenforced_primary_key(["id"]).await.unwrap();
+        table
+            .create_index(
+                &["vec"],
+                Index::IvfPq(
+                    IvfPqIndexBuilder::default()
+                        .distance_type(crate::DistanceType::Cosine)
+                        .num_partitions(1)
+                        .num_sub_vectors(1),
+                ),
+            )
+            .name("vec_cosine".to_string())
+            .execute()
+            .await
+            .unwrap();
+        table
+            .set_lsm_write_spec(
+                LsmWriteSpec::unsharded().with_maintained_indexes(vec!["vec_cosine".to_string()]),
+            )
+            .await
+            .unwrap();
+
+        let base_only = table
+            .query()
+            .nearest_to(query.as_slice())
+            .unwrap()
+            .limit(1)
+            .use_lsm(false)
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        let (base_id, public_distance) = first_result(&base_only);
+
+        let lsm = table
+            .query()
+            .nearest_to(query.as_slice())
+            .unwrap()
+            .limit(1)
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        let (lsm_id, lsm_distance) = first_result(&lsm);
+        assert_eq!(lsm_id, base_id);
+        assert!(
+            (lsm_distance - public_distance).abs() < 1e-5,
+            "LSM cosine distance {lsm_distance} did not use the public scale {public_distance}"
+        );
+
+        // Add an exact memtable result whose distance lies between the public ANN
+        // score and its doubled internal score. Correctly normalized plans still
+        // rank the ANN row first; mixed units would incorrectly rank this row first.
+        assert!(public_distance > 0.0 && public_distance < 4.0 / 3.0);
+        let memtable_distance = public_distance * 1.5;
+        let cosine_similarity = 1.0 - memtable_distance;
+        let mut orthogonal = normalized_vector(&mut state);
+        let projection = orthogonal
+            .iter()
+            .zip(&query)
+            .map(|(left, right)| left * right)
+            .sum::<f32>();
+        for (value, query_value) in orthogonal.iter_mut().zip(&query) {
+            *value -= projection * query_value;
+        }
+        let norm = orthogonal
+            .iter()
+            .map(|value| value * value)
+            .sum::<f32>()
+            .sqrt();
+        orthogonal.iter_mut().for_each(|value| *value /= norm);
+        let sine = (1.0 - cosine_similarity * cosine_similarity).sqrt();
+        let memtable_vector = query
+            .iter()
+            .zip(&orthogonal)
+            .map(|(query_value, orthogonal_value)| {
+                cosine_similarity * query_value + sine * orthogonal_value
+            })
+            .collect::<Vec<_>>();
+
+        let mut merge = table.merge_insert(&[]);
+        merge
+            .when_matched_update_all(None)
+            .when_not_matched_insert_all();
+        let memtable = make_batch(vec![(N as i64, memtable_vector)]);
+        merge
+            .execute(Box::new(RecordBatchIterator::new(
+                vec![Ok(memtable)],
+                schema,
+            )))
+            .await
+            .unwrap();
+
+        let mixed = table
+            .query()
+            .nearest_to(query.as_slice())
+            .unwrap()
+            .limit(1)
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        let (mixed_id, mixed_distance) = first_result(&mixed);
+        assert_eq!(
+            mixed_id, base_id,
+            "mixed LSM tiers must compare ANN and exact distances in public units"
+        );
+        assert!((mixed_distance - public_distance).abs() < 1e-5);
+    }
 }

rust/lancedb/src/table/query.rs

@@ -1,7 +1,10 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
-use std::sync::Arc;
+use std::{
+    collections::{HashSet, VecDeque},
+    sync::Arc,
+};
 
 mod lsm;
 
@@ -17,15 +20,23 @@ use arrow::array::{AsArray, FixedSizeListBuilder, Float32Builder};
 use arrow::datatypes::{Float32Type, UInt8Type};
 use arrow_array::Array;
 use arrow_schema::{DataType, Schema};
-use datafusion_physical_plan::ExecutionPlan;
+use datafusion_common::{Column, DataFusionError, ScalarValue, SchemaError};
+use datafusion_expr::Operator;
+use datafusion_physical_expr::expressions::{BinaryExpr, Column as PhysicalColumn, Literal};
+use datafusion_physical_plan::PhysicalExpr;
 use datafusion_physical_plan::projection::ProjectionExec;
 use datafusion_physical_plan::repartition::RepartitionExec;
 use datafusion_physical_plan::union::UnionExec;
-use futures::future::try_join_all;
+use datafusion_physical_plan::{ExecutionPlan, with_new_children_if_necessary};
 use lance::dataset::mem_wal::DatasetMemWalExt;
 use lance::dataset::scanner::DatasetRecordBatchStream;
 use lance::dataset::scanner::Scanner;
+use lance::index::DatasetIndexInternalExt;
+use lance::io::exec::ANNIvfSubIndexExec;
 use lance_datafusion::exec::{analyze_plan as lance_analyze_plan, execute_plan};
+use lance_index::metrics::NoOpMetricsCollector;
+use lance_index::vector::{DIST_COL, quantizer::QuantizationType};
+use lance_linalg::distance::DistanceType as LanceDistanceType;
 use lance_namespace::LanceNamespace;
 use lance_namespace::models::{
     QueryTableRequest as NsQueryTableRequest, QueryTableRequestColumns,
@@ -45,6 +56,22 @@ impl AnyQuery {
             Self::VectorQuery(query) => &query.base,
         }
     }
+
+    fn base_mut(&mut self) -> &mut QueryRequest {
+        match self {
+            Self::Query(query) => query,
+            Self::VectorQuery(query) => &mut query.base,
+        }
+    }
+
+    /// Canonicalize any raw SQL filter immediately before backend dispatch.
+    pub(crate) fn canonicalized(&self) -> Result<Self> {
+        let mut query = self.clone();
+        if let Some(QueryFilter::Sql(predicate)) = &mut query.base_mut().filter {
+            *predicate = crate::expr::canonicalize_sql_predicate(predicate)?;
+        }
+        Ok(query)
+    }
 }
 
 //Decide between namespace or local
@@ -53,15 +80,16 @@ pub async fn execute_query(
     query: &AnyQuery,
     options: QueryExecutionOptions,
 ) -> Result<DatasetRecordBatchStream> {
+    let query = query.canonicalized()?;
     // QueryTable pushdown runs the query server-side, but only on the main
     // branch: the namespace request carries no branch yet, so a branch handle
     // must fall through to local execution.
-    if can_execute_namespace_query(table, query).await?
+    if can_execute_namespace_query(table, &query).await?
         && let Some(ref namespace_client) = table.namespace_client
     {
-        return execute_namespace_query(table, namespace_client.clone(), query, options).await;
+        return execute_namespace_query(table, namespace_client.clone(), &query, options).await;
     }
-    execute_generic_query(table, query, options).await
+    execute_generic_query(table, &query, options).await
 }
 
 async fn can_execute_namespace_query(table: &NativeTable, query: &AnyQuery) -> Result<bool> {
@@ -136,9 +164,10 @@ pub async fn create_plan(
     query: &AnyQuery,
     options: QueryExecutionOptions,
 ) -> Result<Arc<dyn ExecutionPlan>> {
+    let query = query.canonicalized()?;
     let query = match query {
-        AnyQuery::VectorQuery(query) => query.clone(),
-        AnyQuery::Query(query) => VectorQueryRequest::from_plain_query(query.clone()),
+        AnyQuery::VectorQuery(query) => query,
+        AnyQuery::Query(query) => VectorQueryRequest::from_plain_query(query),
     };
     query.base.check_filter()?;
 
@@ -170,26 +199,48 @@ pub async fn create_plan(
     let mut column = query.column.clone();
 
     let mut query_vector = query.query_vector.first().cloned();
+    let mut is_batch_query = false;
     if query.query_vector.len() > 1 {
         if column.is_none() {
             // Infer a vector column with the same dimension of the query vector.
-            let arrow_schema = Schema::from(ds_ref.schema());
+            let arrow_schema = Schema::from(schema);
             column = Some(default_vector_column(
                 &arrow_schema,
                 Some(query.query_vector[0].len() as i32),
             )?);
         }
         let vector_field = schema.field(column.as_ref().unwrap()).unwrap();
-        if let DataType::List(_) = vector_field.data_type() {
-            // Multivector handling: concatenate into FixedSizeList<FixedSizeList<_>>
+        let (_, element_type) =
+            lance::index::vector::utils::get_vector_type(schema, column.as_ref().unwrap())?;
+        let is_binary = matches!(element_type, DataType::UInt8);
+        if matches!(vector_field.data_type(), DataType::List(_))
+            || (query.base.offset.unwrap_or(0) == 0 && !is_binary)
+        {
+            // Lance distinguishes these cases from the vector column type: a
+            // list-like query against a List column is one multivector query,
+            // while the same query against a FixedSizeList column is a batch of
+            // independent queries. The batch path shares a single flat scan and
+            // bounds retained candidate data instead of running one scan per
+            // query vector.
             let vectors = query
                 .query_vector
                 .iter()
                 .map(|arr| arr.as_ref())
                 .collect::<Vec<_>>();
             let dim = vectors[0].len();
+            if let Some((query_index, actual_dim)) = vectors
+                .iter()
+                .enumerate()
+                .find_map(|(index, vector)| (vector.len() != dim).then_some((index, vector.len())))
+            {
+                return Err(Error::InvalidInput {
+                    message: format!(
+                        "query vector at index {query_index} has dimension {actual_dim}, expected {dim}"
+                    ),
+                });
+            }
             let mut fsl_builder = FixedSizeListBuilder::with_capacity(
-                Float32Builder::with_capacity(dim),
+                Float32Builder::with_capacity(dim * vectors.len()),
                 dim as i32,
                 vectors.len(),
             );
@@ -200,8 +251,12 @@ pub async fn create_plan(
                 fsl_builder.append(true);
             }
             query_vector = Some(Arc::new(fsl_builder.finish()));
+            is_batch_query = !matches!(vector_field.data_type(), DataType::List(_));
         } else {
-            // Multiple query vectors: create a plan for each and union them
+            // Lance's batch path has no per-query offset, and its binary path
+            // requires primitive UInt8 queries rather than a fixed-size list.
+            // Keep the prior plan shape for these cases so offsets are applied
+            // per query and binary query vectors retain their primitive shape.
             let query_vecs = query.query_vector.clone();
             let plan_futures = query_vecs
                 .into_iter()
@@ -214,7 +269,7 @@ pub async fn create_plan(
                     }
                 })
                 .collect::<Vec<_>>();
-            let plans = try_join_all(plan_futures).await?;
+            let plans = futures::future::try_join_all(plan_futures).await?;
             return create_multi_vector_plan(plans);
         }
     }
@@ -225,7 +280,7 @@ pub async fn create_plan(
         let column = if let Some(col) = column {
             col
         } else {
-            let arrow_schema = Schema::from(ds_ref.schema());
+            let arrow_schema = Schema::from(schema);
             default_vector_column(&arrow_schema, Some(query_vector.len() as i32))?
         };
 
@@ -251,10 +306,14 @@ pub async fn create_plan(
         }
     }
 
-    scanner.limit(
-        query.base.limit.map(|limit| limit as i64),
-        query.base.offset.map(|offset| offset as i64),
-    )?;
+    // For a batch query, `nearest` already applies k to each query vector.
+    // Adding Scanner's global limit would truncate the combined result to k rows.
+    if !is_batch_query {
+        scanner.limit(
+            query.base.limit.map(|limit| limit as i64),
+            query.base.offset.map(|offset| offset as i64),
+        )?;
+    }
 
     if let Some(ef) = query.ef {
         scanner.ef(ef);
@@ -327,11 +386,299 @@ pub async fn create_plan(
         scanner.order_by(Some(order_by.clone()))?;
     }
 
-    Ok(scanner.create_plan().await?)
+    let mut plan = scanner
+        .create_plan()
+        .await
+        .map_err(|error| enrich_lance_field_not_found(error, schema))?;
+    let normalized_l2_indices = normalized_l2_ann_indices(plan.as_ref()).await?;
+    if !normalized_l2_indices.is_empty() {
+        // Rebuild only the affected ANN nodes with internal normalized squared-L2
+        // bounds. Exact branches keep the public cosine bounds from `plan`.
+        let internal_plan = if query.lower_bound.is_some() || query.upper_bound.is_some() {
+            scanner.distance_range(
+                query.lower_bound.map(|bound| bound / COSINE_ANN_SCALE),
+                query.upper_bound.map(|bound| bound / COSINE_ANN_SCALE),
+            );
+            scanner
+                .create_plan()
+                .await
+                .map_err(|error| enrich_lance_field_not_found(error, schema))?
+        } else {
+            plan.clone()
+        };
+        plan = normalize_ann_branches(plan, internal_plan, &normalized_l2_indices)?;
+    }
+
+    Ok(plan)
+}
+
+/// Replace DataFusion's top-level field candidates with qualified leaf paths.
+///
+/// DataFusion resolves nested fields but its `FieldNotFound` error only lists the
+/// top-level Arrow fields. This makes a missing leaf look unavailable even when it
+/// exists below a struct. Keep every other Lance/DataFusion error unchanged and
+/// enrich only this one schema error at the LanceDB query boundary.
+fn enrich_lance_field_not_found(
+    error: lance::Error,
+    schema: &lance_core::datatypes::Schema,
+) -> Error {
+    let Some(field) = find_missing_field(&error) else {
+        return error.into();
+    };
+    field_not_found_error(field, &Schema::from(schema))
+}
+
+fn field_not_found_diagnostic(
+    error: &(dyn std::error::Error + 'static),
+    schema: &Schema,
+) -> Option<Error> {
+    let field = find_missing_field(error)?;
+    Some(field_not_found_error(field, schema))
+}
+
+fn field_not_found_error(field: &Column, schema: &Schema) -> Error {
+    let valid_fields = leaf_field_paths(schema);
+    let mut message = format!("Schema error: No field named {}", field.quoted_flat_name());
+    if !valid_fields.is_empty() {
+        message.push_str(". Valid fields are ");
+        message.push_str(&valid_fields.join(", "));
+    }
+    message.push('.');
+
+    Error::InvalidInput { message }
+}
+
+fn find_missing_field<'a>(error: &'a (dyn std::error::Error + 'static)) -> Option<&'a Column> {
+    if let Some(DataFusionError::SchemaError(schema_error, _)) =
+        error.downcast_ref::<DataFusionError>()
+        && let SchemaError::FieldNotFound { field, .. } = schema_error.as_ref()
+    {
+        return Some(field);
+    }
+
+    error.source().and_then(find_missing_field)
+}
+
+fn leaf_field_paths(schema: &Schema) -> Vec<String> {
+    fn format_segment(segment: &str) -> String {
+        // Quote every segment instead of maintaining a SQL keyword list. Bare
+        // lowercase names such as `true` can be parsed as expressions rather
+        // than identifiers, while backticks preserve all field names in both
+        // local SQL parsers.
+        format!("`{}`", segment.replace('`', "``"))
+    }
+
+    fn visit(fields: &arrow_schema::Fields, path: &mut Vec<String>, paths: &mut Vec<String>) {
+        for field in fields {
+            // Neither local planner can address an empty field-path segment,
+            // even when it is backtick-quoted. Do not advertise leaves beneath
+            // such a segment as valid filter fields.
+            if field.name().is_empty() {
+                continue;
+            }
+            path.push(field.name().clone());
+            match field.data_type() {
+                DataType::Struct(children) if !children.is_empty() => {
+                    visit(children, path, paths);
+                }
+                _ => {
+                    paths.push(
+                        path.iter()
+                            .map(|segment| format_segment(segment))
+                            .collect::<Vec<_>>()
+                            .join("."),
+                    );
+                }
+            }
+            path.pop();
+        }
+    }
+
+    let mut paths = Vec::new();
+    visit(schema.fields(), &mut Vec::new(), &mut paths);
+    paths
 }
 
 //Helper functions below
 
+const COSINE_ANN_SCALE: f32 = 0.5;
+
+/// Find ANN index segments whose scores use normalized squared L2 for cosine search.
+///
+/// Cosine PQ/SQ/RQ indices normalize their vectors and use squared L2 internally.  This
+/// preserves ranking, but squared L2 over unit vectors is twice the cosine distance.  Flat
+/// cosine indices calculate cosine directly, so they are not included.
+async fn normalized_l2_ann_indices(plan: &dyn ExecutionPlan) -> Result<HashSet<String>> {
+    let mut ann_plans = Vec::new();
+    find_ann_plans(plan, &mut ann_plans);
+
+    let mut checked = HashSet::new();
+    let mut normalized_l2 = HashSet::new();
+    for ann in ann_plans {
+        if ann.query().metric_type != Some(LanceDistanceType::Cosine) {
+            continue;
+        }
+        for index in ann.indices() {
+            let uuid = index.uuid.to_string();
+            if !checked.insert(uuid.clone()) {
+                continue;
+            }
+            let vector_index = ann
+                .dataset()
+                .open_vector_index(&ann.query().column, &index.uuid, &NoOpMetricsCollector)
+                .await?;
+            let (_, quantization_type) = vector_index.sub_index_type();
+            if matches!(
+                quantization_type,
+                QuantizationType::Product | QuantizationType::Scalar | QuantizationType::Rabit
+            ) {
+                normalized_l2.insert(uuid);
+            }
+        }
+    }
+    Ok(normalized_l2)
+}
+
+/// Normalize affected ANN outputs before their parent plan nodes consume them.
+///
+/// This is used by planners that do not support distance ranges, such as the MemWAL
+/// LSM planner. The standard scanner path rebuilds a second plan when it also needs
+/// to translate range bounds, then calls [`normalize_ann_branches`] directly.
+pub(super) async fn normalize_cosine_ann_branches(
+    plan: Arc<dyn ExecutionPlan>,
+) -> Result<Arc<dyn ExecutionPlan>> {
+    let normalized_l2_indices = normalized_l2_ann_indices(plan.as_ref()).await?;
+    if normalized_l2_indices.is_empty() {
+        return Ok(plan);
+    }
+    normalize_ann_branches(plan.clone(), plan, &normalized_l2_indices)
+}
+
+fn find_ann_plans<'a>(plan: &'a dyn ExecutionPlan, ann_plans: &mut Vec<&'a ANNIvfSubIndexExec>) {
+    if let Some(ann) = plan.downcast_ref::<ANNIvfSubIndexExec>() {
+        ann_plans.push(ann);
+    }
+    for child in plan.children() {
+        find_ann_plans(child.as_ref(), ann_plans);
+    }
+}
+
+fn collect_ann_plans(
+    plan: &Arc<dyn ExecutionPlan>,
+    ann_plans: &mut VecDeque<Arc<dyn ExecutionPlan>>,
+) {
+    if plan.downcast_ref::<ANNIvfSubIndexExec>().is_some() {
+        ann_plans.push_back(plan.clone());
+        return;
+    }
+    for child in plan.children() {
+        collect_ann_plans(child, ann_plans);
+    }
+}
+
+/// Replace normalized-L2 ANN nodes with equivalent nodes that use internal bounds, then
+/// convert their output to the public cosine scale before any generic plan node consumes it.
+fn normalize_ann_branches(
+    public_plan: Arc<dyn ExecutionPlan>,
+    internal_plan: Arc<dyn ExecutionPlan>,
+    normalized_l2_indices: &HashSet<String>,
+) -> Result<Arc<dyn ExecutionPlan>> {
+    let mut internal_ann_plans = VecDeque::new();
+    collect_ann_plans(&internal_plan, &mut internal_ann_plans);
+    let normalized =
+        replace_ann_branches(public_plan, &mut internal_ann_plans, normalized_l2_indices)?;
+    if !internal_ann_plans.is_empty() {
+        return Err(Error::Runtime {
+            message: "internal and public vector plans contained different ANN branches"
+                .to_string(),
+        });
+    }
+    Ok(normalized)
+}
+
+fn replace_ann_branches(
+    public_plan: Arc<dyn ExecutionPlan>,
+    internal_ann_plans: &mut VecDeque<Arc<dyn ExecutionPlan>>,
+    normalized_l2_indices: &HashSet<String>,
+) -> Result<Arc<dyn ExecutionPlan>> {
+    if let Some(public_ann) = public_plan.downcast_ref::<ANNIvfSubIndexExec>() {
+        let internal_plan = internal_ann_plans
+            .pop_front()
+            .ok_or_else(|| Error::Runtime {
+                message: "internal vector plan was missing an ANN branch".to_string(),
+            })?;
+        let internal_ann = internal_plan
+            .downcast_ref::<ANNIvfSubIndexExec>()
+            .expect("collected only ANN plans");
+        let same_indices = public_ann
+            .indices()
+            .iter()
+            .map(|index| &index.uuid)
+            .eq(internal_ann.indices().iter().map(|index| &index.uuid));
+        if public_ann.query().column != internal_ann.query().column
+            || public_ann.query().metric_type != internal_ann.query().metric_type
+            || !same_indices
+        {
+            return Err(Error::Runtime {
+                message: "internal and public vector plans had mismatched ANN branches".to_string(),
+            });
+        }
+
+        let normalized_count = public_ann
+            .indices()
+            .iter()
+            .filter(|index| normalized_l2_indices.contains(&index.uuid.to_string()))
+            .count();
+        if normalized_count == 0 {
+            return Ok(public_plan);
+        }
+        if normalized_count != public_ann.indices().len() {
+            return Err(Error::Runtime {
+                message: "one ANN branch mixed public and normalized-L2 distance scales"
+                    .to_string(),
+            });
+        }
+        return scale_distance_column(internal_plan, COSINE_ANN_SCALE);
+    }
+
+    let children = public_plan
+        .children()
+        .into_iter()
+        .cloned()
+        .map(|child| replace_ann_branches(child, internal_ann_plans, normalized_l2_indices))
+        .collect::<Result<Vec<_>>>()?;
+    Ok(with_new_children_if_necessary(public_plan, children)?)
+}
+
+fn scale_distance_column(
+    plan: Arc<dyn ExecutionPlan>,
+    scale: f32,
+) -> Result<Arc<dyn ExecutionPlan>> {
+    let schema = plan.schema();
+    if schema.column_with_name(DIST_COL).is_none() {
+        return Ok(plan);
+    }
+
+    let expressions: Vec<(Arc<dyn PhysicalExpr>, String)> = schema
+        .fields()
+        .iter()
+        .enumerate()
+        .map(|(index, field)| {
+            let column: Arc<dyn PhysicalExpr> = Arc::new(PhysicalColumn::new(field.name(), index));
+            let expression = if field.name() == DIST_COL {
+                let scale: Arc<dyn PhysicalExpr> =
+                    Arc::new(Literal::new(ScalarValue::Float32(Some(scale))));
+                Arc::new(BinaryExpr::new(column, Operator::Multiply, scale))
+                    as Arc<dyn PhysicalExpr>
+            } else {
+                column
+            };
+            (expression, field.name().clone())
+        })
+        .collect();
+    Ok(Arc::new(ProjectionExec::try_new(expressions, plan)?))
+}
+
 // Take many execution plans and map them into a single plan that adds
 // a query_index column and unions them.
 pub(crate) fn create_multi_vector_plan(
@@ -687,7 +1034,10 @@ async fn parse_arrow_ipc_response(bytes: bytes::Bytes) -> Result<DatasetRecordBa
 #[cfg(test)]
 #[allow(deprecated)]
 mod tests {
-    use arrow_array::{ArrayRef, FixedSizeListArray, Float32Array};
+    use arrow_array::{
+        ArrayRef, FixedSizeListArray, Float32Array, Int32Array, RecordBatch, StringArray,
+        StructArray,
+    };
     use futures::TryStreamExt;
     use lance_arrow::FixedSizeListArrayExt;
     use std::sync::{
@@ -696,7 +1046,7 @@ mod tests {
     };
 
     use super::*;
-    use crate::query::{QueryExecutionOptions, QueryRequest};
+    use crate::query::{ExecutableQuery, QueryBase, QueryExecutionOptions, QueryRequest};
     use crate::table::BaseTable;
 
     fn fixed_size_list_array(values: Vec<f32>, dimension: i32) -> FixedSizeListArray {
@@ -837,7 +1187,6 @@ mod tests {
     async fn test_execute_query_local_routing() {
         use crate::connect;
         use crate::table::query::execute_query;
-        use arrow_array::{Int32Array, RecordBatch};
         use arrow_schema::{DataType, Field, Schema};
 
         let conn = connect("memory://").execute().await.unwrap();
@@ -877,6 +1226,164 @@ mod tests {
         assert_eq!(count, 2); // 4 and 5
     }
 
+    #[tokio::test]
+    async fn test_missing_filter_field_lists_nested_fields_in_local_planners() {
+        use crate::connect;
+        use arrow_schema::{DataType, Field, Schema};
+
+        let conn = connect("memory://").execute().await.unwrap();
+        let metadata = Arc::new(StructArray::from(vec![
+            (
+                Arc::new(Field::new("year", DataType::Int32, false)),
+                Arc::new(Int32Array::from(vec![2024])) as ArrayRef,
+            ),
+            (
+                Arc::new(Field::new("genre", DataType::Utf8, false)),
+                Arc::new(StringArray::from(vec!["fiction"])) as ArrayRef,
+            ),
+            (
+                Arc::new(Field::new("Title", DataType::Int32, false)),
+                Arc::new(Int32Array::from(vec![7])) as ArrayRef,
+            ),
+            (
+                Arc::new(Field::new("true", DataType::Int32, false)),
+                Arc::new(Int32Array::from(vec![8])) as ArrayRef,
+            ),
+            (
+                Arc::new(Field::new("", DataType::Int32, false)),
+                Arc::new(Int32Array::from(vec![10])) as ArrayRef,
+            ),
+        ]));
+        let vector = Arc::new(fixed_size_list_array(vec![0.0, 1.0], 2));
+        let schema = Arc::new(Schema::new(vec![
+            Field::new("id", DataType::Int32, false),
+            Field::new("vector", vector.data_type().clone(), false),
+            Field::new("content", DataType::Utf8, false),
+            Field::new("metadata", metadata.data_type().clone(), false),
+        ]));
+        let batch = RecordBatch::try_new(
+            schema,
+            vec![
+                Arc::new(Int32Array::from(vec![1])),
+                vector,
+                Arc::new(StringArray::from(vec!["example"])),
+                metadata,
+            ],
+        )
+        .unwrap();
+        let table = conn
+            .create_table("nested_error", batch)
+            .execute()
+            .await
+            .unwrap();
+
+        let error = table
+            .query()
+            .only_if("year = 2024")
+            .execute()
+            .await
+            .err()
+            .expect("query should reject the unqualified nested field");
+        let case_sensitive_path = "`metadata`.`Title`";
+        let keyword_path = "`metadata`.`true`";
+        let expected = format!(
+            "No field named year. Valid fields are `id`, `vector`, `content`, `metadata`.`year`, `metadata`.`genre`, {case_sensitive_path}, {keyword_path}."
+        );
+
+        assert!(
+            error.to_string().contains(&expected),
+            "unexpected error: {error}"
+        );
+        for (path, value) in [(case_sensitive_path, 7), (keyword_path, 8)] {
+            table
+                .query()
+                .only_if(format!("{path} = {value}"))
+                .execute()
+                .await
+                .expect("the path advertised by the diagnostic should be reusable");
+        }
+
+        table.set_unenforced_primary_key(["id"]).await.unwrap();
+        table
+            .set_lsm_write_spec(crate::table::LsmWriteSpec::unsharded())
+            .await
+            .unwrap();
+        let lsm_error = table
+            .query()
+            .only_if("year = 2024")
+            .execute()
+            .await
+            .err()
+            .expect("LSM query should reject the unqualified nested field");
+
+        assert!(
+            lsm_error.to_string().contains(&expected),
+            "unexpected LSM error: {lsm_error}"
+        );
+        for (path, value) in [(case_sensitive_path, 7), (keyword_path, 8)] {
+            table
+                .query()
+                .only_if(format!("{path} = {value}"))
+                .execute()
+                .await
+                .expect("the path advertised by the diagnostic should be reusable in LSM queries");
+        }
+    }
+
+    #[test]
+    fn test_leaf_field_paths_preserve_arbitrary_depth() {
+        use arrow_schema::{DataType, Field, Schema};
+
+        fn nested_field(path: &[&str]) -> Field {
+            let mut segments = path.iter().rev();
+            let mut field = Field::new(
+                *segments.next().expect("path must have a leaf"),
+                DataType::Int32,
+                false,
+            );
+            for segment in segments {
+                field = Field::new(*segment, DataType::Struct(vec![field].into()), false);
+            }
+            field
+        }
+
+        let schema = Schema::new(vec![
+            nested_field(&["a", "b", "c", "d", "e"]),
+            nested_field(&["metadata", "child.with.dot"]),
+            nested_field(&["metadata", "Title"]),
+            nested_field(&["metadata", "123child"]),
+            nested_field(&["metadata", "child`tick"]),
+            nested_field(&["metadata", ""]),
+            nested_field(&["", "child"]),
+        ]);
+
+        assert_eq!(
+            leaf_field_paths(&schema),
+            vec![
+                "`a`.`b`.`c`.`d`.`e`",
+                "`metadata`.`child.with.dot`",
+                "`metadata`.`Title`",
+                "`metadata`.`123child`",
+                "`metadata`.`child``tick`",
+            ]
+        );
+
+        let source = DataFusionError::SchemaError(
+            Box::new(SchemaError::FieldNotFound {
+                field: Box::new(Column::from_name("missing")),
+                valid_fields: Vec::new(),
+            }),
+            Box::new(None),
+        );
+        let error = field_not_found_diagnostic(&source, &schema).unwrap();
+        assert!(
+            error.to_string().contains(
+                "Valid fields are `a`.`b`.`c`.`d`.`e`, `metadata`.`child.with.dot`, `metadata`.`Title`, `metadata`.`123child`, `metadata`.`child``tick`"
+            ),
+            "unexpected error: {error}"
+        );
+    }
+
     #[derive(Debug, Default)]
     struct CountingNamespaceClient {
         query_table_calls: AtomicUsize,
@@ -1088,7 +1595,7 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn test_create_plan_multivector_structure() {
+    async fn test_create_plan_batch_vector_uses_shared_scan() {
         use arrow_array::{Float32Array, RecordBatch};
         use arrow_schema::{DataType, Field, Schema};
         use datafusion_physical_plan::display::DisplayableExecutionPlan;
@@ -1115,11 +1622,18 @@ mod tests {
             .unwrap();
         let native_table = table.as_native().unwrap();
 
-        // This triggers the "create_multi_vector_plan" logic branch
+        // A batch of vectors against a fixed-size vector column should use
+        // Lance's native batch KNN path instead of independent scan plans.
         let q1 = Arc::new(Float32Array::from(vec![1.0, 2.0]));
         let q2 = Arc::new(Float32Array::from(vec![3.0, 4.0]));
 
         let req = VectorQueryRequest {
+            base: QueryRequest {
+                filter: Some(QueryFilter::Sql("id >= 0".to_string())),
+                limit: Some(1),
+                select: Select::Columns(vec!["id".to_string()]),
+                ..Default::default()
+            },
             column: Some("vector".to_string()),
             query_vector: vec![q1, q2],
             ..Default::default()
@@ -1136,22 +1650,220 @@ mod tests {
             .indent(true)
             .to_string();
 
-        // We expect a RepartitionExec wrapping a UnionExec
         assert!(
-            display.contains("RepartitionExec"),
-            "Plan should include Repartitioning"
+            display.contains("KNNVectorDistance: queries=2"),
+            "plan should use native batch KNN, got:\n{display}"
         );
         assert!(
-            display.contains("UnionExec"),
-            "Plan should include a Union of multiple searches"
+            !display.contains("UnionExec"),
+            "flat batch KNN should share one scan, got:\n{display}"
         );
-        // We expect the projection to add the 'query_index' column (logic inside multi_vector_plan)
         assert!(
             display.contains("query_index"),
-            "Plan should add query_index column"
+            "plan should add query_index column, got:\n{display}"
         );
     }
 
+    #[tokio::test]
+    async fn test_cosine_pq_distance_uses_public_cosine_scale() {
+        use arrow_array::{Int32Array, RecordBatch, types::Float32Type};
+        use arrow_schema::{DataType, Field, Schema};
+
+        use crate::connect;
+        use crate::index::{Index, vector::IvfPqIndexBuilder};
+
+        fn normalized_vector(state: &mut u64, dimension: usize) -> Vec<f32> {
+            let mut vector = (0..dimension)
+                .map(|_| {
+                    *state = state
+                        .wrapping_mul(6_364_136_223_846_793_005)
+                        .wrapping_add(1);
+                    ((*state >> 32) as u32 as f32 / u32::MAX as f32) * 2.0 - 1.0
+                })
+                .collect::<Vec<_>>();
+            let norm = vector.iter().map(|value| value * value).sum::<f32>().sqrt();
+            vector.iter_mut().for_each(|value| *value /= norm);
+            vector
+        }
+
+        fn distances(batches: &[RecordBatch]) -> Vec<f32> {
+            batches
+                .iter()
+                .flat_map(|batch| {
+                    batch[DIST_COL]
+                        .as_primitive::<Float32Type>()
+                        .values()
+                        .to_vec()
+                })
+                .collect()
+        }
+
+        let conn = connect("memory://").execute().await.unwrap();
+        let dimension = 8;
+        let num_rows = 256;
+        let mut state = 42;
+        let values = (0..num_rows)
+            .flat_map(|_| normalized_vector(&mut state, dimension))
+            .collect::<Vec<_>>();
+        let query_vector = normalized_vector(&mut state, dimension);
+        let vectors = Arc::new(fixed_size_list_array(values, dimension as i32));
+        let schema = Arc::new(Schema::new(vec![
+            Field::new("id", DataType::Int32, false),
+            Field::new("vector", vectors.data_type().clone(), false),
+        ]));
+        let batch = RecordBatch::try_new(
+            schema.clone(),
+            vec![Arc::new(Int32Array::from_iter_values(0..num_rows)), vectors],
+        )
+        .unwrap();
+        let table = conn
+            .create_table("test_cosine_pq_distance", batch)
+            .execute()
+            .await
+            .unwrap();
+        table
+            .create_index(
+                &["vector"],
+                Index::IvfPq(
+                    IvfPqIndexBuilder::default()
+                        .distance_type(crate::DistanceType::Cosine)
+                        .num_partitions(1)
+                        .num_sub_vectors(1),
+                ),
+            )
+            .execute()
+            .await
+            .unwrap();
+
+        let approximate = table
+            .vector_search(query_vector.as_slice())
+            .unwrap()
+            .limit(5)
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        let refined = table
+            .vector_search(query_vector.as_slice())
+            .unwrap()
+            .limit(5)
+            .refine_factor(1)
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        let approximate_distances = distances(&approximate);
+        let refined_distances = distances(&refined);
+        assert_eq!(approximate_distances.len(), refined_distances.len());
+        for (approximate, refined) in approximate_distances.iter().zip(&refined_distances) {
+            assert!(
+                (approximate - refined).abs() < 1e-5,
+                "approximate cosine distance {approximate} did not use the public scale; refined distance was {refined}"
+            );
+        }
+
+        // Distance range bounds are public cosine distances too. Lance applies them to
+        // internal ANN scores, so the planner must translate the bounds before execution.
+        let nearest = approximate_distances[0];
+        let ranged = table
+            .vector_search(query_vector.as_slice())
+            .unwrap()
+            .limit(1)
+            .distance_range(Some(nearest - 1e-5), Some(nearest + 1e-5))
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        let ranged_distances = distances(&ranged);
+        assert_eq!(ranged_distances.len(), 1);
+        assert!((ranged_distances[0] - nearest).abs() < 1e-5);
+
+        let refined_ranged = table
+            .vector_search(query_vector.as_slice())
+            .unwrap()
+            .limit(1)
+            .refine_factor(1)
+            .distance_range(None, Some(nearest + 1e-5))
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        assert_eq!(
+            distances(&refined_ranged).len(),
+            1,
+            "refinement must not apply public cosine bounds to internal ANN scores"
+        );
+
+        let aliased = table
+            .vector_search(query_vector.as_slice())
+            .unwrap()
+            .limit(1)
+            .select(Select::dynamic(&[("aliased_distance", "_distance")]))
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        let batch = &aliased[0];
+        let aliased_distance = batch["aliased_distance"]
+            .as_primitive::<Float32Type>()
+            .value(0);
+        let public_distance = batch[DIST_COL].as_primitive::<Float32Type>().value(0);
+        assert!(
+            (aliased_distance - public_distance).abs() < 1e-5,
+            "distance aliases and auto-projected distances must use the same public scale"
+        );
+
+        // Appended rows take an exact fallback branch. Its public range filter must stay
+        // independent of the translated ANN bounds before both branches are merged.
+        let mut orthogonal = normalized_vector(&mut state, dimension);
+        let projection = orthogonal
+            .iter()
+            .zip(&query_vector)
+            .map(|(left, right)| left * right)
+            .sum::<f32>();
+        for (value, query_value) in orthogonal.iter_mut().zip(&query_vector) {
+            *value -= projection * query_value;
+        }
+        let norm = orthogonal
+            .iter()
+            .map(|value| value * value)
+            .sum::<f32>()
+            .sqrt();
+        orthogonal.iter_mut().for_each(|value| *value /= norm);
+        let appended_vectors = Arc::new(fixed_size_list_array(orthogonal, dimension as i32));
+        let appended = RecordBatch::try_new(
+            schema,
+            vec![Arc::new(Int32Array::from(vec![num_rows])), appended_vectors],
+        )
+        .unwrap();
+        table.add(appended).execute().await.unwrap();
+
+        let mixed = table
+            .vector_search(query_vector.as_slice())
+            .unwrap()
+            .limit(5)
+            .distance_range(None, Some(nearest + 1e-5))
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+        let mixed_distances = distances(&mixed);
+        assert_eq!(mixed_distances.len(), 1);
+        assert!((mixed_distances[0] - nearest).abs() < 1e-5);
+    }
+
     #[tokio::test]
     async fn test_create_plan_applies_approx_mode_to_ann_query() {
         use arrow_array::RecordBatch;

rust/lancedb/src/table/query/lsm.rs

@@ -27,6 +27,8 @@ use std::sync::Arc;
 
 use arrow_array::Array;
 use arrow_schema::{DataType, Schema as ArrowSchema};
+use datafusion::common::{DataFusionError, ToDFSchema};
+use datafusion::prelude::SessionContext;
 use datafusion_physical_plan::expressions::Column;
 use datafusion_physical_plan::projection::ProjectionExec;
 use datafusion_physical_plan::{ExecutionPlan, PhysicalExpr};
@@ -128,6 +130,10 @@ pub(super) async fn create_lsm_plan(
         .await?
     };
 
+    // Normalize cosine ANN arms before LSM merge and sort nodes compare their
+    // distances with exact SSTable and memtable arms.
+    let plan = super::normalize_cosine_ann_branches(plan).await?;
+
     // Lance appends the primary-key columns internally for dedup and keeps them in
     // the output; drop the ones the user did not request so the projection matches.
     restore_projection(plan, &query, &pk_columns)
@@ -391,7 +397,21 @@ fn base_scanner(
     }
     if let Some(filter) = &query.base.filter {
         scanner = match filter {
-            QueryFilter::Sql(sql) => scanner.filter(sql)?,
+            QueryFilter::Sql(sql) => {
+                // Parse here instead of inside `LsmScanner::filter` so the typed
+                // DataFusion `FieldNotFound` error is still available for the
+                // same nested-field enrichment used by the ordinary scanner.
+                let schema = ArrowSchema::from(dataset.schema());
+                let df_schema = schema.clone().to_dfschema().map_err(|error| {
+                    enrich_filter_error(error, &schema, "Failed to create DFSchema")
+                })?;
+                let expr = SessionContext::new()
+                    .parse_sql_expr(sql, &df_schema)
+                    .map_err(|error| {
+                        enrich_filter_error(error, &schema, "Failed to parse filter expression")
+                    })?;
+                scanner.filter_expr(expr)
+            }
             QueryFilter::Datafusion(expr) => scanner.filter_expr(expr.clone()),
             QueryFilter::Substrait(_) => {
                 return Err(Error::NotSupported {
@@ -403,6 +423,12 @@ fn base_scanner(
     Ok(scanner)
 }
 
+fn enrich_filter_error(error: DataFusionError, schema: &ArrowSchema, context: &str) -> Error {
+    super::field_not_found_diagnostic(&error, schema).unwrap_or_else(|| Error::InvalidInput {
+        message: format!("{context}: {error}"),
+    })
+}
+
 /// Plain scan: filter / projection / limit over base ∪ SSTables ∪ in-memory.
 /// The plain scan applies limit and offset inside the planner.
 async fn plain_plan(

rust/lancedb/src/table/schema_evolution.rs

@@ -55,7 +55,9 @@ pub struct DropColumnsResult {
 pub struct FieldMetadataUpdate {
     /// Dot-separated path to the field (e.g. `"embedding"` or `"address.zip"`).
     pub path: String,
-    /// Keys to set (`Some`) or delete (`None`).
+    /// Keys to set (`Some`) or delete (`None`). See
+    /// [`Table::update_field_metadata`](crate::Table::update_field_metadata) for
+    /// the conventional `lancedb:*` keys.
     pub metadata: HashMap<String, Option<String>>,
     /// If `true`, replace the field's entire metadata map instead of merging.
     pub replace: bool,

rust/lancedb/src/table/update.rs

@@ -62,22 +62,33 @@ impl UpdateBuilder {
     }
 
     /// Executes the update operation.
-    pub async fn execute(self) -> Result<UpdateResult> {
+    pub async fn execute(mut self) -> Result<UpdateResult> {
         if self.columns.is_empty() {
             Err(Error::InvalidInput {
                 message: "at least one column must be specified in an update operation".to_string(),
             })
         } else {
+            self.canonicalize_filter()?;
             self.parent.clone().update(self).await
         }
     }
+
+    pub(crate) fn canonicalize_filter(&mut self) -> Result<()> {
+        self.filter = self
+            .filter
+            .take()
+            .map(|predicate| crate::expr::canonicalize_sql_predicate(&predicate))
+            .transpose()?;
+        Ok(())
+    }
 }
 
 /// Internal implementation of the update logic
 pub(crate) async fn execute_update(
     table: &NativeTable,
-    update: UpdateBuilder,
+    mut update: UpdateBuilder,
 ) -> Result<UpdateResult> {
+    update.canonicalize_filter()?;
     table.dataset.ensure_mutable()?;
 
     // 1. Snapshot the current dataset