chore: update lance dependency to v12.0.0-beta.18

this PR blob v2 field helpers and reads to the Node SDK.

`blob()` marks a field as blob v2 and lets you set the storage
thresholds. Inputs can be bytes, a URI, or a data/uri struct.

Queries return descriptors. `fetchBlobs()` reads the bytes by row ID,
and `fetchBlobFiles()` gives you lazy handles for full or range reads.
`blobColumns()` lists the blob fields, including nested ones.

Fetch uses the table’s current checkout. It preserves order, duplicates,
and nulls. Holding row IDs across compaction still requires stable row
IDs.

```javascript
const db = await connect("./data");
const video = await readFile("clip.mp4");
const table = await db.createTable(
  "videos",
  [{ id: 1n, video }],
  {
    schema: new Schema([
      new Field("id", new Int64()),
      blob("video"),
    ]),
  },
);
const rows = await table.query().select(["id"]).withRowId().toArray();
const rowIds = rows.map((row) => row._rowid as bigint);
const bytes = await table.fetchBlobs("video", rowIds);
const [handle] = await table.fetchBlobFiles("video", rowIds);
const header = await handle!.readRange(0n, 65536n);

```

### Testing

- cover input validation, thresholds, nested fields, fetch ordering,
nulls, and range reads.

.bumpversion.toml

@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "0.39.0-beta.2"
+current_version = "0.39.0-beta.6"
 parse = """(?x)
     (?P<major>0|[1-9]\\d*)\\.
     (?P<minor>0|[1-9]\\d*)\\.

.github/workflows/typos.yml

@@ -0,0 +1,20 @@
+name: Typo checker
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+
+permissions:
+  contents: read
+
+jobs:
+  run:
+    name: Spell Check with Typos
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v6
+
+      - name: Check spelling of the entire repository
+        uses: crate-ci/typos@6802cc60d4e7f78b9d5454f6cf3935c042d5e1e3 # v1.26.0

.pre-commit-config.yaml

@@ -10,6 +10,10 @@ repos:
     rev: v0.9.9
     hooks:
       - id: ruff
+  - repo: https://github.com/crate-ci/typos
+    rev: v1.26.0
+    hooks:
+      - id: typos
   # - repo: https://github.com/RobertCraigie/pyright-python
   #   rev: v1.1.395
   #   hooks:

.typos.toml

@@ -0,0 +1,19 @@
+[default]
+extend-ignore-re = ["(?Rm)^.*(#|//)\\s*spellchecker:disable-line$"]
+
+[default.extend-words]
+# Azure Kubernetes Service, mentioned in rust/lancedb/src/remote/oauth.rs.
+AKS = "AKS"
+# RabitQ is the name of a vector quantization algorithm, not a typo of "Rabbit".
+Rabit = "Rabit"
+# `VarBuilder::from_mmaped_safetensors` is the real (if oddly-spelled) name of
+# the candle-core API we call in rust/lancedb/src/embeddings/sentence_transformers.rs.
+mmaped = "mmaped"
+# `WriteableBuffer` is the real name of a type from Python's `_typeshed` stubs,
+# used in python/python/lancedb/_blob.py.
+Writeable = "Writeable"
+
+[files]
+extend-exclude = [
+    "*_THIRD_PARTY_LICENSES.*",
+]

Cargo.toml

@@ -13,20 +13,20 @@ categories = ["database-implementations"]
 rust-version = "1.91.0"
 
 [workspace.dependencies]
-lance = { "version" = "=12.0.0-beta.11", default-features = false, "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-core = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-datagen = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-file = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-io = { "version" = "=12.0.0-beta.11", default-features = false, "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-index = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-linalg = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-namespace = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-namespace-impls = { "version" = "=12.0.0-beta.11", default-features = false, "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-table = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-testing = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-datafusion = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-encoding = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
-lance-arrow = { "version" = "=12.0.0-beta.11", "tag" = "v12.0.0-beta.11", "git" = "https://github.com/lance-format/lance.git" }
+lance = { "version" = "=12.0.0-beta.18", default-features = false, "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-core = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-datagen = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-file = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-io = { "version" = "=12.0.0-beta.18", default-features = false, "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-index = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-linalg = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace-impls = { "version" = "=12.0.0-beta.18", default-features = false, "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-table = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-testing = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-datafusion = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-encoding = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
+lance-arrow = { "version" = "=12.0.0-beta.18", "tag" = "v12.0.0-beta.18", "git" = "https://github.com/lance-format/lance.git" }
 lancedb = { path = "rust/lancedb", default-features = false }
 ahash = "0.8"
 # Note that this one does not include pyarrow
@@ -60,7 +60,7 @@ log = "0.4"
 metrics = "0.24"
 metrics-util = "0.19"
 moka = { version = "0.12", features = ["future"] }
-object_store = "0.13.2"
+object_store = "0.14.1"
 pin-project = "1.0.7"
 rand = "0.9"
 snafu = "0.8"

docs/openapi.yml

@@ -155,7 +155,7 @@ paths:
                 vector:
                   type: FixedSizeList
                   description: |
-                    The targetted vector to search for. Required.
+                    The targeted vector to search for. Required.
                 vector_column:
                   type: string
                   description: |

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>0.39.0-beta.2</version>
+    <version>0.39.0-beta.6</version>
 </dependency>
 ```
 

docs/src/js/classes/BlobFile.md

@@ -0,0 +1,62 @@
+[**@lancedb/lancedb**](../README.md) • **Docs**
+
+***
+
+[@lancedb/lancedb](../globals.md) / BlobFile
+
+# Class: BlobFile
+
+A lazy handle to blob bytes. Create one with [Table.fetchBlobFiles](Table.md#fetchblobfiles).
+
+## Methods
+
+### read()
+
+```ts
+read(): Promise<Buffer>
+```
+
+Reads from the cursor to the end and advances the cursor.
+
+A second call returns an empty buffer. [BlobFile.readRange](BlobFile.md#readrange) does
+not move the cursor.
+
+#### Returns
+
+`Promise`&lt;`Buffer`&gt;
+
+***
+
+### readRange()
+
+```ts
+readRange(start, end): Promise<Buffer>
+```
+
+Reads the half-open byte range `[start, end)`.
+
+Fails when `end` is past the blob size. Does not move the cursor.
+
+#### Parameters
+
+* **start**: `bigint`
+
+* **end**: `bigint`
+
+#### Returns
+
+`Promise`&lt;`Buffer`&gt;
+
+***
+
+### size()
+
+```ts
+size(): bigint
+```
+
+Returns the blob size in bytes.
+
+#### Returns
+
+`bigint`

docs/src/js/classes/MergeInsertBuilder.md

@@ -141,7 +141,7 @@ Currently this causes multiple copies of the row to be created
 but that behavior is subject to change.
 
 An optional condition may be specified.  If it is, then only
-matched rows that satisfy the condtion will be updated.  Any
+matched rows that satisfy the condition will be updated.  Any
 rows that do not satisfy the condition will be left as they
 are.  Failing to satisfy the condition does not cause a
 "matched row" to become a "not matched" row.

docs/src/js/classes/Table.md

@@ -137,6 +137,20 @@ containing the new version number of the table after altering the columns.
 
 ***
 
+### blobColumns()
+
+```ts
+abstract blobColumns(): Promise<string[]>
+```
+
+Blob v2 columns, including nested dotted paths.
+
+#### Returns
+
+`Promise`&lt;`string`[]&gt;
+
+***
+
 ### branches()
 
 ```ts
@@ -499,6 +513,54 @@ Drop an index from the table.
 
 ***
 
+### fetchBlobFiles()
+
+```ts
+abstract fetchBlobFiles(column, rowIds): Promise<(null | BlobFile)[]>
+```
+
+Opens lazy blob handles for `column` at the given row IDs using the
+table's current checkout.
+
+Preserves input order, duplicates, and nulls. Use this for large payloads.
+See [Table.fetchBlobs](Table.md#fetchblobs) for row-ID validity across versions.
+
+#### Parameters
+
+* **column**: `string`
+
+* **rowIds**: readonly (`number` \| `bigint`)[]
+
+#### Returns
+
+`Promise`&lt;(`null` \| [`BlobFile`](BlobFile.md))[]&gt;
+
+***
+
+### fetchBlobs()
+
+```ts
+abstract fetchBlobs(column, rowIds): Promise<(null | Buffer)[]>
+```
+
+Bytes for `column` at row IDs from [Query.withRowId](Query.md#withrowid).
+
+Reads the table's current checkout. IDs from another version can fail after
+compaction unless stable row ids are enabled. Results keep input order and
+duplicates. Null blobs are `null`. Empty blobs are empty buffers.
+
+#### Parameters
+
+* **column**: `string`
+
+* **rowIds**: readonly (`number` \| `bigint`)[]
+
+#### Returns
+
+`Promise`&lt;(`null` \| `Buffer`)[]&gt;
+
+***
+
 ### flushLsm()
 
 ```ts
@@ -1266,7 +1328,7 @@ value is 0")
 Note: if your condition is something like "some_id_column == 7" and
 you are updating many rows (with different ids) then you will get
 better performance with a single [`merge_insert`] call instead of
-repeatedly calilng this method.
+repeatedly calling this method.
 
 ##### Parameters
 

docs/src/js/functions/blob.md

@@ -0,0 +1,55 @@
+[**@lancedb/lancedb**](../README.md) • **Docs**
+
+***
+
+[@lancedb/lancedb](../globals.md) / blob
+
+# Function: blob()
+
+```ts
+function blob(name, options): Field
+```
+
+Declares a `lance.blob.v2` column.
+
+Query results are descriptors, not payload bytes. Use [Table.fetchBlobs](../classes/Table.md#fetchblobs)
+or [Table.fetchBlobFiles](../classes/Table.md#fetchblobfiles) to read bytes.
+
+## Parameters
+
+* **name**: `string`
+
+* **options**: [`BlobOptions`](../type-aliases/BlobOptions.md) = `{}`
+
+## Returns
+
+`Field`
+
+## Example
+
+```ts
+import { readFile } from "node:fs/promises";
+import { Field, Int64, Schema } from "apache-arrow";
+import { blob, connect } from "@lancedb/lancedb";
+
+const db = await connect("./data");
+const video = await readFile("clip.mp4");
+const table = await db.createTable(
+  "videos",
+  [{ id: 1n, video }],
+  {
+    schema: new Schema([
+      new Field("id", new Int64()),
+      blob("video"),
+    ]),
+  },
+);
+
+const rows = await table.query().select(["id"]).withRowId().toArray();
+const rowIds = rows.map((row) => row._rowid as bigint);
+const bytes = await table.fetchBlobs("video", rowIds);
+
+const [handle] = await table.fetchBlobFiles("video", rowIds);
+const size = handle!.size();
+const header = await handle!.readRange(0n, size < 65536n ? size : 65536n);
+```

docs/src/js/functions/isBlobField.md

@@ -0,0 +1,22 @@
+[**@lancedb/lancedb**](../README.md) • **Docs**
+
+***
+
+[@lancedb/lancedb](../globals.md) / isBlobField
+
+# Function: isBlobField()
+
+```ts
+function isBlobField(field): boolean
+```
+
+Checks for the `lance.blob.v2` extension marker. Does not validate the
+field's storage type.
+
+## Parameters
+
+* **field**: `Field`<`any`>
+
+## Returns
+
+`boolean`

docs/src/js/globals.md

@@ -19,6 +19,7 @@
 ## Classes
 
 - [AutoQuery](classes/AutoQuery.md)
+- [BlobFile](classes/BlobFile.md)
 - [BooleanQuery](classes/BooleanQuery.md)
 - [BoostQuery](classes/BoostQuery.md)
 - [BranchContents](classes/BranchContents.md)
@@ -143,6 +144,7 @@
 
 - [AnalyzePlanDistributedMetrics](type-aliases/AnalyzePlanDistributedMetrics.md)
 - [BaseTokenizer](type-aliases/BaseTokenizer.md)
+- [BlobOptions](type-aliases/BlobOptions.md)
 - [Data](type-aliases/Data.md)
 - [DataLike](type-aliases/DataLike.md)
 - [FieldLike](type-aliases/FieldLike.md)
@@ -158,9 +160,11 @@
 ## Functions
 
 - [RecordBatchIterator](functions/RecordBatchIterator.md)
+- [blob](functions/blob.md)
 - [connect](functions/connect.md)
 - [connectNamespace](functions/connectNamespace.md)
 - [instrumentLanceDbMetrics](functions/instrumentLanceDbMetrics.md)
+- [isBlobField](functions/isBlobField.md)
 - [makeArrowTable](functions/makeArrowTable.md)
 - [packBits](functions/packBits.md)
 - [permutationBuilder](functions/permutationBuilder.md)

docs/src/js/interfaces/HnswPqOptions.md

@@ -118,7 +118,7 @@ Number of sub-vectors of PQ.
 This value controls how much the vector is compressed during the quantization step.
 The more sub vectors there are the less the vector is compressed.  The default is
 the dimension of the vector divided by 16.  If the dimension is not evenly divisible
-by 16 we use the dimension divded by 8.
+by 16 we use the dimension divided by 8.
 
 The above two cases are highly preferred.  Having 8 or 16 values per subvector allows
 us to use efficient SIMD instructions.

docs/src/js/interfaces/IndexOptions.md

@@ -16,7 +16,7 @@ optional config: Index;
 
 Advanced index configuration
 
-This option allows you to specify a specfic index to create and also
+This option allows you to specify a specific index to create and also
 allows you to pass in configuration for training the index.
 
 See the static methods on Index for details on the various index types.

docs/src/js/interfaces/IvfPqOptions.md

@@ -112,7 +112,7 @@ Number of sub-vectors of PQ.
 This value controls how much the vector is compressed during the quantization step.
 The more sub vectors there are the less the vector is compressed.  The default is
 the dimension of the vector divided by 16.  If the dimension is not evenly divisible
-by 16 we use the dimension divded by 8.
+by 16 we use the dimension divided by 8.
 
 The above two cases are highly preferred.  Having 8 or 16 values per subvector allows
 us to use efficient SIMD instructions.

docs/src/js/type-aliases/BlobOptions.md

@@ -0,0 +1,48 @@
+[**@lancedb/lancedb**](../README.md) • **Docs**
+
+***
+
+[@lancedb/lancedb](../globals.md) / BlobOptions
+
+# Type Alias: BlobOptions
+
+```ts
+type BlobOptions: object;
+```
+
+## Type declaration
+
+### dedicatedSizeThreshold?
+
+```ts
+optional dedicatedSizeThreshold: number;
+```
+
+Max payload bytes stored in a packed sidecar before a dedicated file. Must
+be a positive safe integer.
+
+### inlineSizeThreshold?
+
+```ts
+optional inlineSizeThreshold: number;
+```
+
+Max payload bytes kept inline in the data file. Zero is allowed. Must be a
+safe integer.
+
+### nullable?
+
+```ts
+optional nullable: boolean;
+```
+
+Defaults to true.
+
+### packFileSizeThreshold?
+
+```ts
+optional packFileSizeThreshold: number;
+```
+
+Max bytes in one packed sidecar before starting another. Must be a positive
+safe integer.

java/lancedb-core/pom.xml

@@ -8,7 +8,7 @@
     <parent>
       <groupId>com.lancedb</groupId>
       <artifactId>lancedb-parent</artifactId>
-      <version>0.39.0-beta.2</version>
+      <version>0.39.0-beta.6</version>
       <relativePath>../pom.xml</relativePath>
     </parent>
 

java/pom.xml

@@ -6,7 +6,7 @@
 
     <groupId>com.lancedb</groupId>
     <artifactId>lancedb-parent</artifactId>
-    <version>0.39.0-beta.2</version>
+    <version>0.39.0-beta.6</version>
     <packaging>pom</packaging>
     <name>${project.artifactId}</name>
     <description>LanceDB Java SDK Parent POM</description>
@@ -28,7 +28,7 @@
     <properties>
         <project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
         <arrow.version>15.0.0</arrow.version>
-        <lance-core.version>12.0.0-beta.11</lance-core.version>
+        <lance-core.version>12.0.0-beta.18</lance-core.version>
         <spotless.skip>false</spotless.skip>
         <spotless.version>2.30.0</spotless.version>
         <spotless.java.googlejavaformat.version>1.7</spotless.java.googlejavaformat.version>

nodejs/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
 name = "lancedb-nodejs"
 edition.workspace = true
-version = "0.39.0-beta.2"
+version = "0.39.0-beta.6"
 publish = false
 license.workspace = true
 description.workspace = true

nodejs/__test__/blob.test.ts

@@ -0,0 +1,185 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+import { Field, Int64, List, Schema, Struct, Utf8 } from "apache-arrow";
+import { makeArrowTable } from "../lancedb/arrow";
+import { BlobFile, blob, coerceBlobValue, isBlobField } from "../lancedb/blob";
+
+describe("blob()", () => {
+  it("marks the field as lance.blob.v2", () => {
+    const field = blob("image", { nullable: false });
+    expect(field.nullable).toBe(false);
+    expect(isBlobField(field)).toBe(true);
+    expect(field.metadata.get("ARROW:extension:name")).toBe("lance.blob.v2");
+  });
+
+  it("writes encoding thresholds as field metadata", () => {
+    const field = blob("video", {
+      inlineSizeThreshold: 1024,
+      dedicatedSizeThreshold: 2 * 1024 * 1024,
+      packFileSizeThreshold: 64 * 1024 * 1024,
+    });
+    expect(
+      field.metadata.get("lance-encoding:blob-inline-size-threshold"),
+    ).toBe("1024");
+    expect(
+      field.metadata.get("lance-encoding:blob-dedicated-size-threshold"),
+    ).toBe(String(2 * 1024 * 1024));
+    expect(
+      field.metadata.get("lance-encoding:blob-pack-file-size-threshold"),
+    ).toBe(String(64 * 1024 * 1024));
+  });
+
+  it("rejects invalid thresholds", () => {
+    expect(() => blob("image", { inlineSizeThreshold: -1 })).toThrow(
+      /inlineSizeThreshold must be non-negative/,
+    );
+    expect(() => blob("image", { dedicatedSizeThreshold: 0 })).toThrow(
+      /dedicatedSizeThreshold must be positive/,
+    );
+    expect(() => blob("image", { packFileSizeThreshold: 1.5 })).toThrow(
+      /packFileSizeThreshold must be a safe integer/,
+    );
+    expect(() =>
+      blob("image", { dedicatedSizeThreshold: Number.MAX_SAFE_INTEGER + 1 }),
+    ).toThrow(/dedicatedSizeThreshold must be a safe integer/);
+  });
+});
+
+describe("coerceBlobValue", () => {
+  it.each([
+    ["Buffer", Buffer.from("x"), { data: Buffer.from("x"), uri: null }],
+    [
+      "Uint8Array",
+      new Uint8Array([120]),
+      { data: new Uint8Array([120]), uri: null },
+    ],
+    ["URI string", "s3://bucket/key", { data: null, uri: "s3://bucket/key" }],
+    [
+      "data struct",
+      { data: Buffer.from("y") },
+      { data: Buffer.from("y"), uri: null },
+    ],
+    [
+      "uri struct",
+      { uri: "s3://bucket/key" },
+      { data: null, uri: "s3://bucket/key" },
+    ],
+    ["null", null, null],
+  ])("accepts %s", (_name, input, expected) => {
+    expect(coerceBlobValue(input)).toEqual(expected);
+  });
+
+  it.each([
+    ["empty URI", "", /uri cannot be empty/],
+    ["object without data or uri", { position: 0 }, /data' or 'uri/],
+    [
+      "Int16Array",
+      new Int16Array([1]),
+      /Blob data must be Buffer or Uint8Array/,
+    ],
+    [
+      "both data and uri",
+      { data: Buffer.from("y"), uri: "s3://bucket/key" },
+      /exactly one of 'data' or 'uri'/,
+    ],
+    [
+      "neither data nor uri",
+      { data: null, uri: null },
+      /exactly one of 'data' or 'uri'/,
+    ],
+  ])("rejects %s", (_name, input, message) => {
+    expect(() => coerceBlobValue(input)).toThrow(message);
+  });
+});
+
+describe("BlobFile", () => {
+  it("rejects constructing BlobFile without a native handle", () => {
+    expect(() => new (BlobFile as unknown as { new (): BlobFile })()).toThrow(
+      /fetchBlobFiles/,
+    );
+  });
+});
+
+describe("makeArrowTable blob columns", () => {
+  it("coerces Buffer input onto a blob field", () => {
+    const schema = new Schema([
+      new Field("id", new Int64(), true),
+      blob("image"),
+    ]);
+    const table = makeArrowTable([{ id: 1n, image: Buffer.from("hello") }], {
+      schema,
+    });
+    expect(isBlobField(table.schema.fields[1])).toBe(true);
+    const image = table.getChild("image")!;
+    expect(image.nullCount).toBe(0);
+    expect(image.getChild("uri")!.get(0)).toBeNull();
+    expect(image.getChild("data")!.nullCount).toBe(0);
+    expect(Buffer.from(image.getChild("data")!.get(0)!).toString()).toBe(
+      "hello",
+    );
+  });
+
+  it("coerces Buffer elements inside a list and keeps null slots", () => {
+    const schema = new Schema([
+      new Field("id", new Int64(), true),
+      new Field("images", new List(blob("image")), true),
+    ]);
+    const table = makeArrowTable(
+      [
+        { id: 1n, images: [Buffer.from("a"), Buffer.from("bb")] },
+        { id: 2n, images: null },
+        { id: 3n, images: [Buffer.from("c"), null] },
+        { id: 4n, images: [] },
+      ],
+      { schema },
+    );
+    const images = table.getChild("images")!;
+    expect(images.nullCount).toBe(1);
+    const rows = images.toArray();
+    expect(rows[1]).toBeNull();
+    expect(Array.from(rows[3] as Iterable<unknown>)).toHaveLength(0);
+    const first = Array.from(rows[0] as Iterable<{ data: Uint8Array | null }>);
+    expect(Buffer.from(first[0].data!).toString()).toBe("a");
+    expect(Buffer.from(first[1].data!).toString()).toBe("bb");
+    const third = Array.from(
+      rows[2] as Iterable<{ data: Uint8Array | null } | null>,
+    );
+    expect(Buffer.from(third[0]!.data!).toString()).toBe("c");
+    expect(third[1]).toBeNull();
+  });
+
+  it("coerces Buffer fields inside list structs", () => {
+    const schema = new Schema([
+      new Field("id", new Int64(), true),
+      new Field(
+        "items",
+        new List(
+          new Field(
+            "item",
+            new Struct([new Field("name", new Utf8(), true), blob("image")]),
+            true,
+          ),
+        ),
+        true,
+      ),
+    ]);
+    const table = makeArrowTable(
+      [
+        {
+          id: 1n,
+          items: [{ name: "one", image: Buffer.from("alpha") }],
+        },
+      ],
+      { schema },
+    );
+    const items = Array.from(
+      table.getChild("items")!.toArray()[0] as Iterable<{
+        name: string;
+        image: { data: Uint8Array | null };
+      }>,
+    );
+    expect(items[0].name).toBe("one");
+    expect(Buffer.from(items[0].image.data!).toString()).toBe("alpha");
+  });
+});

nodejs/__test__/table.test.ts

@@ -18,6 +18,7 @@ import {
   Query,
   Table,
   VectorQuery,
+  blob,
   connect,
   tokenize,
 } from "../lancedb";
@@ -281,7 +282,7 @@ describe.each([arrow15, arrow16, arrow17, arrow18])(
         numIndices: 0,
         numRows: 3,
         // Full on-disk size of the two data files, footers and metadata included.
-        totalBytes: 684,
+        totalBytes: 550,
       });
 
       // Index files count toward totalBytes too (only deletion files and
@@ -289,7 +290,7 @@ describe.each([arrow15, arrow16, arrow17, arrow18])(
       await table.createIndex("id", { config: Index.btree() });
       const statsWithIndex = await table.stats();
       expect(statsWithIndex.numIndices).toBe(1);
-      expect(statsWithIndex.totalBytes).toBeGreaterThan(684);
+      expect(statsWithIndex.totalBytes).toBeGreaterThan(550);
     });
 
     it("should overwrite data if asked", async () => {
@@ -2401,6 +2402,276 @@ describe("when dealing with versioning", () => {
   });
 });
 
+describe("when dealing with blob columns", () => {
+  let tmpDir: tmp.DirResult;
+  beforeEach(() => {
+    tmpDir = tmp.dirSync({ unsafeCleanup: true });
+  });
+  afterEach(() => {
+    tmpDir.removeCallback();
+  });
+
+  it("discovers blob columns", async () => {
+    const { table } = await openBlobTable();
+    expect(await table.blobColumns()).toEqual(["image"]);
+  });
+
+  it("preserves order, duplicates, and nulls", async () => {
+    const { table, rowIds } = await openBlobTable();
+    const [alphaId, betaId, nullId] = rowIds;
+    const bytes = await table.fetchBlobs("image", [
+      betaId,
+      alphaId,
+      betaId,
+      nullId,
+    ]);
+    expect(bytes.map((b) => (b == null ? null : b.toString()))).toEqual([
+      "beta",
+      "alpha",
+      "beta",
+      null,
+    ]);
+    const files = await table.fetchBlobFiles("image", [
+      betaId,
+      nullId,
+      alphaId,
+    ]);
+    expect(files.map((f) => f == null)).toEqual([false, true, false]);
+  });
+
+  it("reads full blob contents", async () => {
+    const { table, rowIds, alpha, beta } = await openBlobTable();
+    const bytes = await table.fetchBlobs("image", rowIds);
+    expect(bytes[0]!.equals(alpha)).toBe(true);
+    expect(bytes[1]!.equals(beta)).toBe(true);
+    const files = await table.fetchBlobFiles("image", rowIds);
+    expect(files[0]!.size()).toBe(BigInt(alpha.length));
+    expect(Buffer.from(await files[0]!.read()).toString()).toBe("alpha");
+    expect(Buffer.from(await files[1]!.read()).toString()).toBe("beta");
+  });
+
+  it("reads a half-open range", async () => {
+    const { table, rowIds } = await openBlobTable();
+    const files = await table.fetchBlobFiles("image", rowIds);
+    expect(Buffer.from(await files[0]!.readRange(0n, 2n)).toString()).toBe(
+      "al",
+    );
+  });
+
+  it("readRange does not move the cursor", async () => {
+    const { table, rowIds, alpha } = await openBlobTable();
+    const [handle] = await table.fetchBlobFiles("image", rowIds);
+    expect((await handle!.readRange(1n, 3n)).toString()).toBe("lp");
+    expect(await handle!.read()).toEqual(alpha);
+    expect(await handle!.read()).toEqual(Buffer.alloc(0));
+  });
+
+  it("fails when readRange end is past the blob size", async () => {
+    const { table, rowIds, alpha } = await openBlobTable();
+    const files = await table.fetchBlobFiles("image", rowIds);
+    await expect(
+      files[0]!.readRange(0n, BigInt(alpha.length + 1)),
+    ).rejects.toThrow(/exceeds blob size/);
+  });
+
+  it("rejects fetchBlobs on a non-blob column", async () => {
+    const { table, rowIds } = await openBlobTable();
+    await expect(table.fetchBlobs("id", rowIds)).rejects.toThrow(/blob/i);
+  });
+
+  it("discovers and fetches nested blob columns", async () => {
+    const db = await connect(tmpDir.name);
+    const schema = new Schema([
+      new Field("id", new Int64(), true),
+      new Field("info", new Struct([blob("image")]), true),
+    ]);
+    const payload = Buffer.from("nested");
+    const table = await db.createTable(
+      "nested_blobs",
+      [{ id: 1n, info: { image: payload } }],
+      { schema },
+    );
+    expect(await table.blobColumns()).toEqual(["info.image"]);
+    const rows = await table.query().withRowId().toArray();
+    const bytes = await table.fetchBlobs("info.image", [
+      rows[0]._rowid as bigint,
+    ]);
+    expect(bytes[0]!.equals(payload)).toBe(true);
+  });
+
+  it("creates and adds list blob columns", async () => {
+    const db = await connect(tmpDir.name);
+    const schema = new Schema([
+      new Field("id", new Int64(), true),
+      new Field("images", new List(blob("image")), true),
+    ]);
+    const alpha = Buffer.from("alpha");
+    const beta = Buffer.from("beta");
+    const gamma = Buffer.from("gamma");
+    const table = await db.createTable(
+      "list_blobs",
+      [{ id: 1n, images: [alpha, beta] }],
+      { schema },
+    );
+    await table.add([
+      { id: 2n, images: null },
+      { id: 3n, images: [gamma, null] },
+      { id: 4n, images: [] },
+    ]);
+    expect(await table.blobColumns()).toEqual(["images.image"]);
+    const rows = await table.query().toArray();
+    const byId = new Map(rows.map((row) => [Number(row.id), row]));
+    expect(descriptorSizes(byId.get(1)!.images)).toEqual([
+      alpha.length,
+      beta.length,
+    ]);
+    expect(byId.get(2)!.images).toBeNull();
+    expect(descriptorSizes(byId.get(3)!.images)).toEqual([gamma.length, null]);
+    expect(Array.from(byId.get(4)!.images as Iterable<unknown>)).toHaveLength(
+      0,
+    );
+  });
+
+  it("creates and adds list struct blob columns", async () => {
+    const db = await connect(tmpDir.name);
+    const schema = new Schema([
+      new Field("id", new Int64(), true),
+      new Field(
+        "items",
+        new List(
+          new Field(
+            "item",
+            new Struct([new Field("name", new Utf8(), true), blob("image")]),
+            true,
+          ),
+        ),
+        true,
+      ),
+    ]);
+    const alpha = Buffer.from("nested-alpha");
+    const beta = Buffer.from("nested-beta");
+    const table = await db.createTable(
+      "list_struct_blobs",
+      [{ id: 1n, items: [{ name: "one", image: alpha }] }],
+      { schema },
+    );
+    await table.add([
+      {
+        id: 2n,
+        items: [
+          { name: "two", image: beta },
+          { name: "three", image: null },
+        ],
+      },
+    ]);
+    const rows = await table.query().toArray();
+    const byId = new Map(rows.map((row) => [Number(row.id), row]));
+    expect(
+      descriptorSizes(
+        Array.from(byId.get(1)!.items as Iterable<{ image: unknown }>).map(
+          (item) => item.image,
+        ),
+      ),
+    ).toEqual([alpha.length]);
+    expect(
+      descriptorSizes(
+        Array.from(byId.get(2)!.items as Iterable<{ image: unknown }>).map(
+          (item) => item.image,
+        ),
+      ),
+    ).toEqual([beta.length, null]);
+  });
+
+  it("rejects blob fields inside a fixed-size list", async () => {
+    const db = await connect(tmpDir.name);
+    const schema = new Schema([
+      new Field("id", new Int64(), true),
+      new Field("frames", new FixedSizeList(2, blob("frame")), true),
+    ]);
+    await expect(
+      db.createTable(
+        "fsl_blobs",
+        [{ id: 1n, frames: [Buffer.from("a"), Buffer.from("b")] }],
+        { schema },
+      ),
+    ).rejects.toThrow(
+      "Blob fields inside FixedSizeList are not supported. Use List instead.",
+    );
+  });
+
+  it("rejects blob fields inside a nested fixed-size list", async () => {
+    const db = await connect(tmpDir.name);
+    const schema = new Schema([
+      new Field("id", new Int64(), true),
+      new Field(
+        "clip",
+        new Struct([
+          new Field("frames", new FixedSizeList(2, blob("frame")), true),
+        ]),
+        true,
+      ),
+    ]);
+    await expect(
+      db.createTable(
+        "nested_fsl_blobs",
+        [
+          {
+            id: 1n,
+            clip: { frames: [Buffer.from("a"), Buffer.from("b")] },
+          },
+        ],
+        { schema },
+      ),
+    ).rejects.toThrow(
+      "Blob fields inside FixedSizeList are not supported. Use List instead.",
+    );
+  });
+
+  it("rejects an Arrow table with blob fields inside a fixed-size list", async () => {
+    const db = await connect(tmpDir.name);
+    const schema = new Schema([
+      new Field("id", new Int64(), true),
+      new Field("frames", new FixedSizeList(2, blob("frame")), true),
+    ]);
+    await expect(
+      db.createTable("fsl_blobs_ipc", new ArrowTable(schema)),
+    ).rejects.toThrow(
+      "Blob fields inside FixedSizeList are not supported. Use List instead.",
+    );
+  });
+
+  function descriptorSizes(values: unknown): (number | null)[] {
+    return Array.from(
+      values as Iterable<{ size?: bigint | number } | null>,
+    ).map((value) => (value == null ? null : Number(value.size)));
+  }
+
+  async function openBlobTable() {
+    const db = await connect(tmpDir.name);
+    const schema = new Schema([
+      new Field("id", new Int64(), true),
+      blob("image"),
+    ]);
+    const alpha = Buffer.from("alpha");
+    const beta = Buffer.from("beta");
+    const table = await db.createTable(
+      "blobs",
+      [
+        { id: 1n, image: alpha },
+        { id: 2n, image: beta },
+        { id: 3n, image: null },
+      ],
+      { schema },
+    );
+    const rows = await table.query().withRowId().toArray();
+    const rowIdById = new Map(
+      rows.map((r) => [Number(r.id), r._rowid as bigint]),
+    );
+    const rowIds = [1, 2, 3].map((id) => rowIdById.get(id)!);
+    return { table, rowIds, alpha, beta };
+  }
+});
+
 describe("when dealing with tags", () => {
   let tmpDir: tmp.DirResult;
   beforeEach(() => {
@@ -3252,7 +3523,7 @@ describe.each([arrow15, arrow16, arrow17, arrow18])(
       const db = await connect(tmpDir.name);
       const data = [
         { text: "fa", vector: [0.1, 0.2, 0.3] },
-        { text: "fo", vector: [0.4, 0.5, 0.6] },
+        { text: "fo", vector: [0.4, 0.5, 0.6] }, // spellchecker:disable-line
         { text: "fob", vector: [0.4, 0.5, 0.6] },
         { text: "focus", vector: [0.4, 0.5, 0.6] },
         { text: "foo", vector: [0.4, 0.5, 0.6] },
@@ -3277,7 +3548,7 @@ describe.each([arrow15, arrow16, arrow17, arrow18])(
       const resultSet = new Set(fuzzyResults.map((r) => r.text));
       expect(resultSet.has("foo")).toBe(true);
       expect(resultSet.has("fob")).toBe(true);
-      expect(resultSet.has("fo")).toBe(true);
+      expect(resultSet.has("fo")).toBe(true); // spellchecker:disable-line
       expect(resultSet.has("food")).toBe(true);
 
       const prefixResults = await table

nodejs/lancedb/arrow.ts

@@ -40,6 +40,7 @@ import {
 } from "apache-arrow";
 import { Buffers } from "apache-arrow/data";
 import { typedArrayToArrowType } from "./arrow_type";
+import { coerceBlobValue, isBlobField } from "./blob";
 import { type EmbeddingFunction } from "./embedding/embedding_function";
 import {
   EmbeddingFunctionConfig,
@@ -430,12 +431,14 @@ export function makeArrowTable(
       throw new Error("A schema must be provided if data is empty");
     } else {
       schema = new Schema(schema.fields, schemaMetadata);
+      validateBlobSchema(schema);
       return new ArrowTable(schema);
     }
   }
 
   let inferredSchema = inferSchema(data, schema, opt);
   inferredSchema = new Schema(inferredSchema.fields, schemaMetadata);
+  validateBlobSchema(inferredSchema);
 
   const finalColumns: Record<string, Vector> = {};
   for (const field of inferredSchema.fields) {
@@ -445,6 +448,35 @@ export function makeArrowTable(
   return new ArrowTable(inferredSchema, finalColumns);
 }
 
+function validateBlobSchema(schema: Schema): void {
+  for (const field of schema.fields) {
+    validateBlobField(field);
+  }
+}
+
+function validateBlobField(field: Field): void {
+  if (
+    isFixedSizeList(field.type) &&
+    containsBlobField(field.type.children[0])
+  ) {
+    throw new Error(
+      "Blob fields inside FixedSizeList are not supported. Use List instead.",
+    );
+  }
+  for (const child of field.type.children ?? []) {
+    validateBlobField(child);
+  }
+}
+
+function containsBlobField(field: Field): boolean {
+  if (isBlobField(field)) {
+    return true;
+  }
+  return (field.type.children ?? []).some((child: Field) =>
+    containsBlobField(child),
+  );
+}
+
 function isObject(value: unknown): value is Record<string, unknown> {
   return (
     typeof value === "object" &&
@@ -480,6 +512,32 @@ function transposeData(
   path: string[] = [],
 ): Vector {
   const valuesPath = [...path, field.name];
+  if (isBlobField(field) && field.type instanceof Struct) {
+    const blobRows = data.map((datum) =>
+      coerceBlobValue(valueAtPath(datum, valuesPath)),
+    );
+    const childVectors = field.type.children.map((child) => {
+      const values = blobRows.map((row) =>
+        row == null ? null : (row[child.name as "data" | "uri"] ?? null),
+      );
+      return makeVector(values, child.type, undefined, child.nullable);
+    });
+    const nullCount = blobRows.filter((row) => row === null).length;
+    const structData = makeData({
+      type: field.type,
+      length: blobRows.length,
+      nullCount,
+      nullBitmap:
+        nullCount > 0
+          ? arrowUtil.packBools(blobRows.map((row) => row !== null))
+          : undefined,
+      children: childVectors.map((v) => v.data[0]),
+    });
+    return arrowMakeVector(structData);
+  }
+  if (isList(field.type) && containsBlobField(field.type.children[0])) {
+    return transposeListData(data, field, valuesPath);
+  }
   const values = data.map((datum) => valueAtPath(datum, valuesPath));
   if (field.type instanceof Struct) {
     const childFields = field.type.children;
@@ -495,7 +553,7 @@ function transposeData(
         nullCount > 0
           ? arrowUtil.packBools(values.map((value) => value !== null))
           : undefined,
-      children: childVectors as unknown as ArrowData<DataType>[],
+      children: childVectors.map((v) => v.data[0]),
     });
     return arrowMakeVector(structData);
   } else {
@@ -503,6 +561,48 @@ function transposeData(
   }
 }
 
+function transposeListData(
+  data: Record<string, unknown>[],
+  field: Field,
+  valuesPath: string[],
+): Vector {
+  const listType = field.type as List;
+  const childField = listType.children[0];
+  const lists = data.map((datum) => valueAtPath(datum, valuesPath));
+  const flattened: Record<string, unknown>[] = [];
+  const validity: boolean[] = [];
+  const offsets: number[] = [0];
+
+  for (const list of lists) {
+    if (list == null) {
+      validity.push(false);
+      offsets.push(flattened.length);
+      continue;
+    }
+    if (!Array.isArray(list)) {
+      throw new Error(`expected an array for list field '${field.name}'`);
+    }
+    validity.push(true);
+    for (const element of list) {
+      flattened.push({ [childField.name]: element });
+    }
+    offsets.push(flattened.length);
+  }
+
+  const childVector = transposeData(flattened, childField, []);
+  const nullCount = validity.filter((valid) => !valid).length;
+  return arrowMakeVector(
+    makeData({
+      type: listType,
+      length: lists.length,
+      nullCount,
+      nullBitmap: nullCount > 0 ? arrowUtil.packBools(validity) : undefined,
+      valueOffsets: Int32Array.from(offsets),
+      child: childVector.data[0],
+    }),
+  );
+}
+
 /**
  * Create an empty Arrow table with the provided schema
  */
@@ -600,7 +700,7 @@ function makeVector(
   }
   if (values.length === 0) {
     throw Error(
-      "makeVector requires at least one value or the type must be specfied",
+      "makeVector requires at least one value or the type must be specified",
     );
   }
   const sampleValue = values.find((val) => val !== null && val !== undefined);
@@ -858,7 +958,7 @@ async function applyEmbeddings<T>(
  * customized by the `embeddingDataType` property of the embedding function.
  *
  * If a schema is provided in `makeTableOptions` then it should include the
- * embedding columns.  If no schema is provded then embedding columns will
+ * embedding columns.  If no schema is provided then embedding columns will
  * be placed at the end of the table, after all of the input columns.
  */
 export async function convertToTable(
@@ -952,6 +1052,7 @@ export async function fromTableToBuffer(
     schema = sanitizeSchema(schema);
   }
   const tableWithEmbeddings = await applyEmbeddings(table, embeddings, schema);
+  validateBlobSchema(tableWithEmbeddings.schema);
   const writer = RecordBatchFileWriter.writeAll(tableWithEmbeddings);
   return Buffer.from(await writer.toUint8Array());
 }

nodejs/lancedb/blob.ts

@@ -0,0 +1,236 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+import { Field, LargeBinary, Struct, Utf8 } from "apache-arrow";
+import { BlobFile as NativeBlobFile } from "./native";
+
+const BLOB_V2_EXTENSION_NAME = "lance.blob.v2";
+
+const INLINE_SIZE_THRESHOLD_KEY = "lance-encoding:blob-inline-size-threshold";
+const DEDICATED_SIZE_THRESHOLD_KEY =
+  "lance-encoding:blob-dedicated-size-threshold";
+const PACK_FILE_SIZE_THRESHOLD_KEY =
+  "lance-encoding:blob-pack-file-size-threshold";
+
+export type BlobInput = {
+  data: Buffer | Uint8Array | null;
+  uri: string | null;
+};
+
+export type BlobOptions = {
+  /** Defaults to true. */
+  nullable?: boolean;
+  /**
+   * Max payload bytes kept inline in the data file. Zero is allowed. Must be a
+   * safe integer.
+   */
+  inlineSizeThreshold?: number;
+  /**
+   * Max payload bytes stored in a packed sidecar before a dedicated file. Must
+   * be a positive safe integer.
+   */
+  dedicatedSizeThreshold?: number;
+  /**
+   * Max bytes in one packed sidecar before starting another. Must be a positive
+   * safe integer.
+   */
+  packFileSizeThreshold?: number;
+};
+
+/**
+ * Declares a `lance.blob.v2` column.
+ *
+ * Query results are descriptors, not payload bytes. Use {@link Table.fetchBlobs}
+ * or {@link Table.fetchBlobFiles} to read bytes.
+ *
+ * @example
+ * ```ts
+ * import { readFile } from "node:fs/promises";
+ * import { Field, Int64, Schema } from "apache-arrow";
+ * import { blob, connect } from "@lancedb/lancedb";
+ *
+ * const db = await connect("./data");
+ * const video = await readFile("clip.mp4");
+ * const table = await db.createTable(
+ *   "videos",
+ *   [{ id: 1n, video }],
+ *   {
+ *     schema: new Schema([
+ *       new Field("id", new Int64()),
+ *       blob("video"),
+ *     ]),
+ *   },
+ * );
+ *
+ * const rows = await table.query().select(["id"]).withRowId().toArray();
+ * const rowIds = rows.map((row) => row._rowid as bigint);
+ * const bytes = await table.fetchBlobs("video", rowIds);
+ *
+ * const [handle] = await table.fetchBlobFiles("video", rowIds);
+ * const size = handle!.size();
+ * const header = await handle!.readRange(0n, size < 65536n ? size : 65536n);
+ * ```
+ */
+export function blob(name: string, options: BlobOptions = {}): Field {
+  const metadata = new Map<string, string>([
+    ["ARROW:extension:name", BLOB_V2_EXTENSION_NAME],
+  ]);
+  setThreshold(
+    metadata,
+    INLINE_SIZE_THRESHOLD_KEY,
+    "inlineSizeThreshold",
+    options.inlineSizeThreshold,
+    0,
+  );
+  setThreshold(
+    metadata,
+    DEDICATED_SIZE_THRESHOLD_KEY,
+    "dedicatedSizeThreshold",
+    options.dedicatedSizeThreshold,
+    1,
+  );
+  setThreshold(
+    metadata,
+    PACK_FILE_SIZE_THRESHOLD_KEY,
+    "packFileSizeThreshold",
+    options.packFileSizeThreshold,
+    1,
+  );
+  return new Field(
+    name,
+    new Struct([
+      new Field("data", new LargeBinary(), true),
+      new Field("uri", new Utf8(), true),
+    ]),
+    options.nullable ?? true,
+    metadata,
+  );
+}
+
+/**
+ * Checks for the `lance.blob.v2` extension marker. Does not validate the
+ * field's storage type.
+ */
+export function isBlobField(field: Field): boolean {
+  return field.metadata?.get("ARROW:extension:name") === BLOB_V2_EXTENSION_NAME;
+}
+
+/**
+ * A lazy handle to blob bytes. Create one with {@link Table.fetchBlobFiles}.
+ *
+ * @hideconstructor
+ */
+export class BlobFile {
+  private readonly inner: NativeBlobFile;
+
+  private constructor(inner: NativeBlobFile) {
+    if (!(inner instanceof NativeBlobFile)) {
+      throw new Error("BlobFile handles come from Table.fetchBlobFiles");
+    }
+    this.inner = inner;
+  }
+
+  /** @ignore */
+  static fromNative(inner: NativeBlobFile): BlobFile {
+    return new BlobFile(inner);
+  }
+
+  /** Returns the blob size in bytes. */
+  size(): bigint {
+    return this.inner.size();
+  }
+
+  /**
+   * Reads from the cursor to the end and advances the cursor.
+   *
+   * A second call returns an empty buffer. {@link BlobFile.readRange} does
+   * not move the cursor.
+   */
+  read(): Promise<Buffer> {
+    return this.inner.read();
+  }
+
+  /**
+   * Reads the half-open byte range `[start, end)`.
+   *
+   * Fails when `end` is past the blob size. Does not move the cursor.
+   */
+  readRange(start: bigint, end: bigint): Promise<Buffer> {
+    return this.inner.readRange(start, end);
+  }
+}
+
+export function coerceBlobValue(value: unknown): BlobInput | null {
+  if (value == null) {
+    return null;
+  }
+  if (isBlobBytes(value)) {
+    return { data: value, uri: null };
+  }
+  if (ArrayBuffer.isView(value)) {
+    throw new Error("Blob data must be Buffer or Uint8Array");
+  }
+  if (typeof value === "string") {
+    if (value === "") {
+      throw new Error("Blob uri cannot be empty");
+    }
+    return { data: null, uri: value };
+  }
+  if (typeof value === "object") {
+    const record = value as Record<string, unknown>;
+    if (!("data" in record) && !("uri" in record)) {
+      throw new Error(
+        "Blob struct values must include a 'data' or 'uri' field",
+      );
+    }
+    const uri = record.uri;
+    if (uri === "") {
+      throw new Error("Blob uri cannot be empty");
+    }
+    if (uri != null && typeof uri !== "string") {
+      throw new Error(`Blob uri must be a string or null, got ${typeof uri}`);
+    }
+    const data = record.data;
+    if (data != null && !isBlobBytes(data)) {
+      throw new Error("Blob data must be Buffer, Uint8Array, or null");
+    }
+    const bytes = (data as Buffer | Uint8Array | null | undefined) ?? null;
+    const uriValue = uri ?? null;
+    if ((bytes == null) === (uriValue == null)) {
+      throw new Error(
+        "Blob struct values must set exactly one of 'data' or 'uri'",
+      );
+    }
+    return { data: bytes, uri: uriValue };
+  }
+  throw new Error(
+    "Blob column values must be Buffer, Uint8Array, a URI string, null, or { data?, uri? }",
+  );
+}
+
+function isBlobBytes(value: unknown): value is Buffer | Uint8Array {
+  return Buffer.isBuffer(value) || value instanceof Uint8Array;
+}
+
+function setThreshold(
+  metadata: Map<string, string>,
+  key: string,
+  optionName: string,
+  value: number | undefined,
+  minimum: number,
+): void {
+  if (value === undefined) {
+    return;
+  }
+  if (!Number.isSafeInteger(value)) {
+    throw new Error(`${optionName} must be a safe integer`);
+  }
+  if (value < minimum) {
+    throw new Error(
+      minimum <= 0
+        ? `${optionName} must be non-negative`
+        : `${optionName} must be positive`,
+    );
+  }
+  metadata.set(key, String(value));
+}

nodejs/lancedb/index.ts

@@ -77,6 +77,9 @@ export {
   VectorColumnOptions,
 } from "./arrow";
 
+export { blob, isBlobField, BlobFile } from "./blob";
+export type { BlobOptions } from "./blob";
+
 export {
   Connection,
   CreateTableOptions,

nodejs/lancedb/indices.ts

@@ -26,7 +26,7 @@ export interface IvfPqOptions {
    * This value controls how much the vector is compressed during the quantization step.
    * The more sub vectors there are the less the vector is compressed.  The default is
    * the dimension of the vector divided by 16.  If the dimension is not evenly divisible
-   * by 16 we use the dimension divded by 8.
+   * by 16 we use the dimension divided by 8.
    *
    * The above two cases are highly preferred.  Having 8 or 16 values per subvector allows
    * us to use efficient SIMD instructions.
@@ -228,7 +228,7 @@ export interface HnswPqOptions {
    * This value controls how much the vector is compressed during the quantization step.
    * The more sub vectors there are the less the vector is compressed.  The default is
    * the dimension of the vector divided by 16.  If the dimension is not evenly divisible
-   * by 16 we use the dimension divded by 8.
+   * by 16 we use the dimension divided by 8.
    *
    * The above two cases are highly preferred.  Having 8 or 16 values per subvector allows
    * us to use efficient SIMD instructions.
@@ -825,7 +825,7 @@ export interface IndexOptions {
   /**
    * Advanced index configuration
    *
-   * This option allows you to specify a specfic index to create and also
+   * This option allows you to specify a specific index to create and also
    * allows you to pass in configuration for training the index.
    *
    * See the static methods on Index for details on the various index types.

nodejs/lancedb/merge.ts

@@ -27,7 +27,7 @@ export class MergeInsertBuilder {
    * but that behavior is subject to change.
    *
    * An optional condition may be specified.  If it is, then only
-   * matched rows that satisfy the condtion will be updated.  Any
+   * matched rows that satisfy the condition will be updated.  Any
    * rows that do not satisfy the condition will be left as they
    * are.  Failing to satisfy the condition does not cause a
    * "matched row" to become a "not matched" row.

nodejs/lancedb/sanitize.ts

@@ -3,7 +3,7 @@
 
 // The utilities in this file help sanitize data from the user's arrow
 // library into the types expected by vectordb's arrow library.  Node
-// generally allows for mulitple versions of the same library (and sometimes
+// generally allows for multiple versions of the same library (and sometimes
 // even multiple copies of the same version) to be installed at the same
 // time.  However, arrow-js uses instanceof which expected that the input
 // comes from the exact same library instance.  This is not always the case

nodejs/lancedb/table.ts

@@ -17,6 +17,7 @@ import {
   tableFromIPC,
 } from "./arrow";
 
+import { BlobFile } from "./blob";
 import { EmbeddingFunctionConfig, getRegistry } from "./embedding/registry";
 import { IndexOptions } from "./indices";
 import { Job } from "./job";
@@ -313,7 +314,7 @@ export abstract class Table {
    * Note: if your condition is something like "some_id_column == 7" and
    * you are updating many rows (with different ids) then you will get
    * better performance with a single [`merge_insert`] call instead of
-   * repeatedly calilng this method.
+   * repeatedly calling this method.
    * @param {Map<string, string> | Record<string, string>} updates - the
    * columns to update
    * @returns {Promise<UpdateResult>} A promise that resolves to an object
@@ -510,6 +511,35 @@ export abstract class Table {
    */
   abstract takeRowIds(rowIds: readonly (bigint | number)[]): TakeQuery;
 
+  /**
+   * Blob v2 columns, including nested dotted paths.
+   */
+  abstract blobColumns(): Promise<string[]>;
+
+  /**
+   * Bytes for `column` at row IDs from {@link Query.withRowId}.
+   *
+   * Reads the table's current checkout. IDs from another version can fail after
+   * compaction unless stable row ids are enabled. Results keep input order and
+   * duplicates. Null blobs are `null`. Empty blobs are empty buffers.
+   */
+  abstract fetchBlobs(
+    column: string,
+    rowIds: readonly (bigint | number)[],
+  ): Promise<(Buffer | null)[]>;
+
+  /**
+   * Opens lazy blob handles for `column` at the given row IDs using the
+   * table's current checkout.
+   *
+   * Preserves input order, duplicates, and nulls. Use this for large payloads.
+   * See {@link Table.fetchBlobs} for row-ID validity across versions.
+   */
+  abstract fetchBlobFiles(
+    column: string,
+    rowIds: readonly (bigint | number)[],
+  ): Promise<(BlobFile | null)[]>;
+
   /**
    * Create a search query to find the nearest neighbors
    * of the given query
@@ -1160,23 +1190,34 @@ export class LocalTable extends Table {
   }
 
   takeRowIds(rowIds: readonly (bigint | number)[]): TakeQuery {
-    const ids = rowIds.map((id) => {
-      if (typeof id === "bigint") {
-        return id;
-      }
-      if (!Number.isInteger(id)) {
-        throw new Error("Row id must be an integer (or bigint)");
-      }
-      if (id < 0) {
-        throw new Error("Row id cannot be negative");
-      }
-      if (!Number.isSafeInteger(id)) {
-        throw new Error("Row id is too large for number; use bigint instead");
-      }
-      return BigInt(id);
-    });
+    return new TakeQuery(this.inner.takeRowIds(rowIdsToBigInts(rowIds)));
+  }
 
-    return new TakeQuery(this.inner.takeRowIds(ids));
+  blobColumns(): Promise<string[]> {
+    return this.inner.blobColumns();
+  }
+
+  async fetchBlobs(
+    column: string,
+    rowIds: readonly (bigint | number)[],
+  ): Promise<(Buffer | null)[]> {
+    const values = await this.inner.fetchBlobs(column, rowIdsToBigInts(rowIds));
+    // N-API Option maps missing values to undefined. Collapse those to null.
+    return values.map((value) => value ?? null);
+  }
+
+  async fetchBlobFiles(
+    column: string,
+    rowIds: readonly (bigint | number)[],
+  ): Promise<(BlobFile | null)[]> {
+    const files = await this.inner.fetchBlobFiles(
+      column,
+      rowIdsToBigInts(rowIds),
+    );
+    // N-API Option maps missing values to undefined. Collapse those to null.
+    return files.map((file) =>
+      file == null ? null : BlobFile.fromNative(file),
+    );
   }
 
   query(): Query {
@@ -1733,3 +1774,21 @@ export class Branches {
     )) as unknown as CherryPickResult;
   }
 }
+
+function rowIdsToBigInts(rowIds: readonly (bigint | number)[]): bigint[] {
+  return rowIds.map((id) => {
+    if (typeof id === "bigint") {
+      return id;
+    }
+    if (!Number.isInteger(id)) {
+      throw new Error("Row id must be an integer (or bigint)");
+    }
+    if (id < 0) {
+      throw new Error("Row id cannot be negative");
+    }
+    if (!Number.isSafeInteger(id)) {
+      throw new Error("Row id is too large for number; use bigint instead");
+    }
+    return BigInt(id);
+  });
+}

nodejs/npm/darwin-arm64/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-darwin-arm64",
-	"version": "0.39.0-beta.2",
+	"version": "0.39.0-beta.6",
 	"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": "0.39.0-beta.2",
+	"version": "0.39.0-beta.6",
 	"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": "0.39.0-beta.2",
+	"version": "0.39.0-beta.6",
 	"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": "0.39.0-beta.2",
+	"version": "0.39.0-beta.6",
 	"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": "0.39.0-beta.2",
+	"version": "0.39.0-beta.6",
 	"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": "0.39.0-beta.2",
+  "version": "0.39.0-beta.6",
   "os": [
     "win32"
   ],

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

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-win32-x64-msvc",
-	"version": "0.39.0-beta.2",
+	"version": "0.39.0-beta.6",
 	"os": ["win32"],
 	"cpu": ["x64"],
 	"main": "lancedb.win32-x64-msvc.node",

nodejs/package.json

@@ -11,7 +11,7 @@
     "ann"
   ],
   "private": false,
-  "version": "0.39.0-beta.2",
+  "version": "0.39.0-beta.6",
   "main": "dist/index.js",
   "exports": {
     ".": "./dist/index.js",

nodejs/pnpm-lock.yaml

@@ -41,7 +41,7 @@ importers:
         version: 3.7.0(@emnapi/core@1.10.0)(@emnapi/runtime@1.11.3)(@types/node@22.7.4)
       '@opentelemetry/sdk-metrics':
         specifier: ^2.10.0
-        version: 2.10.0(@opentelemetry/api@1.9.1)
+        version: 2.11.0(@opentelemetry/api@1.9.1)
       '@types/axios':
         specifier: ^0.14.0
         version: 0.14.4
@@ -80,7 +80,7 @@ importers:
         version: 0.2.7
       ts-jest:
         specifier: ^29.1.2
-        version: 29.4.9(@babel/core@7.29.7)(@jest/transform@29.7.0)(@jest/types@29.6.3)(babel-jest@29.7.0(@babel/core@7.29.7))(jest-util@29.7.0)(jest@29.7.0(@types/node@22.7.4))(typescript@5.5.4)
+        version: 29.4.12(@babel/core@7.29.7)(@jest/transform@29.7.0)(@jest/types@29.6.3)(babel-jest@29.7.0(@babel/core@7.29.7))(jest-util@29.7.0)(jest@29.7.0(@types/node@22.7.4))(typescript@5.5.4)
       typedoc:
         specifier: 0.26.4
         version: 0.26.4(typescript@5.5.4)
@@ -1394,20 +1394,20 @@ packages:
     resolution: {integrity: sha512-gLyJlPHPZYdAk1JENA9LeHejZe1Ti77/pTeFm/nMXmQH/HFZlcS/O2XJB+L8fkbrNSqhdtlvjBVjxwUYanNH5Q==}
     engines: {node: '>=8.0.0'}
 
-  '@opentelemetry/core@2.10.0':
-    resolution: {integrity: sha512-/wNZ8twnEQQA4HoHu22+vcsdru6pWPWxW+7w+FlxT6Id7PE/WIbZmVKkte+PF72e0F2dnImFeHD2syyE1Mw6MQ==}
+  '@opentelemetry/core@2.11.0':
+    resolution: {integrity: sha512-7YP44XH0tV6+Mb54x2YGf84i7yi+31MBZlE8JwvozkxyTvXbSp10X7cI7YE49ChJ3shMJoBmCJF3+1QFBJctGA==}
     engines: {node: ^18.19.0 || >=20.6.0}
     peerDependencies:
       '@opentelemetry/api': '>=1.0.0 <1.10.0'
 
-  '@opentelemetry/resources@2.10.0':
-    resolution: {integrity: sha512-q6MMm2zhggzsHVNbabYwut+a6nbuQQe3URUoxaojM/8K1IBfwwPzvxIjNi2/lI1TFe+fMHMW9MWhrtDLEXEnkA==}
+  '@opentelemetry/resources@2.11.0':
+    resolution: {integrity: sha512-Ie7+8q8MDF4FAEQCKVMTx3ReUvxiIAgIiiW3c9JdmP8+HMcDy20puT+AHjexnExgnbvBxjQ9fjkFDWrikJ2jQA==}
     engines: {node: ^18.19.0 || >=20.6.0}
     peerDependencies:
       '@opentelemetry/api': '>=1.3.0 <1.10.0'
 
-  '@opentelemetry/sdk-metrics@2.10.0':
-    resolution: {integrity: sha512-t6r1VSvXNtSDnPXU1FbZeetJb7yyovHmgu0wRSoftxtE0g2rSNhQZQUy69sRUCL+iioJpX8SN/S6wq6ZtvLySQ==}
+  '@opentelemetry/sdk-metrics@2.11.0':
+    resolution: {integrity: sha512-7GXXcObyHyDUUSG+L+kJoquty01bzm7ivE7+SSgXXJcHuPzGviptxwARmI2c+bnnxjexGQbJnyNlN8HxBP/Y7A==}
     engines: {node: ^18.19.0 || >=20.6.0}
     peerDependencies:
       '@opentelemetry/api': '>=1.9.0 <1.10.0'
@@ -1480,6 +1480,7 @@ packages:
   '@smithy/core@3.24.1':
     resolution: {integrity: sha512-3mT7o4qQyUWttYnVK3A0Z/u3Xha3E81tXn32Tz6vjZiUXhBrkEivpw1hBYfh84iFF9CSzkBU9Y1DJ3Q6RQ231g==}
     engines: {node: '>=18.0.0'}
+    deprecated: Deprecated due to bug in browser bundling instructions https://github.com/smithy-lang/smithy-typescript/issues/2025
 
   '@smithy/credential-provider-imds@4.3.1':
     resolution: {integrity: sha512-0S/acwHnqX4WrjXzhdiDRxsG2s9SC0cpPIK9nZ1R6UOHd+j7uL28+4bHu22urbLk2TVw3fkp6na/+fkUt/pLNQ==}
@@ -3238,8 +3239,8 @@ packages:
     peerDependencies:
       typescript: '>=4.2.0'
 
-  ts-jest@29.4.9:
-    resolution: {integrity: sha512-LTb9496gYPMCqjeDLdPrKuXtncudeV1yRZnF4Wo5l3SFi0RYEnYRNgMrFIdg+FHvfzjCyQk1cLncWVqiSX+EvQ==}
+  ts-jest@29.4.12:
+    resolution: {integrity: sha512-Ov6ClY53Fflh6BGAnY2DlTq1hYDrTycz2PVTXBWFW2CU+9zrEqAp9fWdGXl42EXO5RLSFAcAZ2JFKbP+zBTFfw==}
     engines: {node: ^14.15.0 || ^16.10.0 || ^18.0.0 || >=20.0.0}
     hasBin: true
     peerDependencies:
@@ -5110,22 +5111,22 @@ snapshots:
 
   '@opentelemetry/api@1.9.1': {}
 
-  '@opentelemetry/core@2.10.0(@opentelemetry/api@1.9.1)':
+  '@opentelemetry/core@2.11.0(@opentelemetry/api@1.9.1)':
     dependencies:
       '@opentelemetry/api': 1.9.1
       '@opentelemetry/semantic-conventions': 1.43.0
 
-  '@opentelemetry/resources@2.10.0(@opentelemetry/api@1.9.1)':
+  '@opentelemetry/resources@2.11.0(@opentelemetry/api@1.9.1)':
     dependencies:
       '@opentelemetry/api': 1.9.1
-      '@opentelemetry/core': 2.10.0(@opentelemetry/api@1.9.1)
+      '@opentelemetry/core': 2.11.0(@opentelemetry/api@1.9.1)
       '@opentelemetry/semantic-conventions': 1.43.0
 
-  '@opentelemetry/sdk-metrics@2.10.0(@opentelemetry/api@1.9.1)':
+  '@opentelemetry/sdk-metrics@2.11.0(@opentelemetry/api@1.9.1)':
     dependencies:
       '@opentelemetry/api': 1.9.1
-      '@opentelemetry/core': 2.10.0(@opentelemetry/api@1.9.1)
-      '@opentelemetry/resources': 2.10.0(@opentelemetry/api@1.9.1)
+      '@opentelemetry/core': 2.11.0(@opentelemetry/api@1.9.1)
+      '@opentelemetry/resources': 2.11.0(@opentelemetry/api@1.9.1)
 
   '@opentelemetry/semantic-conventions@1.43.0': {}
 
@@ -5574,7 +5575,7 @@ snapshots:
       globby: 11.1.0
       is-glob: 4.0.3
       minimatch: 9.0.9
-      semver: 7.8.0
+      semver: 7.8.5
       ts-api-utils: 1.4.3(typescript@5.5.4)
     optionalDependencies:
       typescript: 5.5.4
@@ -6426,7 +6427,7 @@ snapshots:
       '@babel/parser': 7.29.3
       '@istanbuljs/schema': 0.1.6
       istanbul-lib-coverage: 3.2.2
-      semver: 7.8.0
+      semver: 7.8.5
     transitivePeerDependencies:
       - supports-color
 
@@ -6705,7 +6706,7 @@ snapshots:
       jest-util: 29.7.0
       natural-compare: 1.4.0
       pretty-format: 29.7.0
-      semver: 7.8.0
+      semver: 7.8.5
     transitivePeerDependencies:
       - supports-color
 
@@ -6828,7 +6829,7 @@ snapshots:
 
   make-dir@4.0.0:
     dependencies:
-      semver: 7.8.0
+      semver: 7.8.5
 
   make-error@1.3.6: {}
 
@@ -7162,8 +7163,7 @@ snapshots:
 
   semver@7.8.0: {}
 
-  semver@7.8.5:
-    optional: true
+  semver@7.8.5: {}
 
   sharp@0.35.4(@types/node@22.7.4):
     dependencies:
@@ -7327,7 +7327,7 @@ snapshots:
     dependencies:
       typescript: 5.5.4
 
-  ts-jest@29.4.9(@babel/core@7.29.7)(@jest/transform@29.7.0)(@jest/types@29.6.3)(babel-jest@29.7.0(@babel/core@7.29.7))(jest-util@29.7.0)(jest@29.7.0(@types/node@22.7.4))(typescript@5.5.4):
+  ts-jest@29.4.12(@babel/core@7.29.7)(@jest/transform@29.7.0)(@jest/types@29.6.3)(babel-jest@29.7.0(@babel/core@7.29.7))(jest-util@29.7.0)(jest@29.7.0(@types/node@22.7.4))(typescript@5.5.4):
     dependencies:
       bs-logger: 0.2.6
       fast-json-stable-stringify: 2.1.0
@@ -7336,7 +7336,7 @@ snapshots:
       json5: 2.2.3
       lodash.memoize: 4.1.2
       make-error: 1.3.6
-      semver: 7.8.0
+      semver: 7.8.5
       type-fest: 4.41.0
       typescript: 5.5.4
       yargs-parser: 21.1.1

nodejs/src/blob.rs

@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+use std::ops::Range;
+use std::sync::Arc;
+
+use arrow_array::{Array, LargeBinaryArray};
+use lancedb::blob::BlobFile as LanceBlobFile;
+use napi::bindgen_prelude::*;
+use napi_derive::napi;
+
+use crate::error::convert_error;
+
+#[napi]
+pub struct BlobFile {
+    inner: Arc<LanceBlobFile>,
+}
+
+impl BlobFile {
+    pub(crate) fn new(inner: LanceBlobFile) -> Self {
+        Self {
+            inner: Arc::new(inner),
+        }
+    }
+}
+
+#[napi]
+impl BlobFile {
+    #[napi]
+    pub fn size(&self) -> BigInt {
+        BigInt::from(self.inner.size())
+    }
+
+    #[napi]
+    pub async fn read(&self) -> napi::Result<Buffer> {
+        let bytes = self.inner.read().await.map_err(|err| convert_error(&err))?;
+        Ok(Buffer::from(bytes.as_ref()))
+    }
+
+    #[napi]
+    pub async fn read_range(&self, start: BigInt, end: BigInt) -> napi::Result<Buffer> {
+        let range = bigint_range(start, end)?;
+        let bytes = self
+            .inner
+            .read_range(range)
+            .await
+            .map_err(|err| convert_error(&err))?;
+        Ok(Buffer::from(bytes.as_ref()))
+    }
+}
+
+fn bigint_range(start: BigInt, end: BigInt) -> napi::Result<Range<u64>> {
+    let start = parse_u64(start, "start")?;
+    let end = parse_u64(end, "end")?;
+    if start > end {
+        return Err(napi::Error::from_reason(format!(
+            "invalid blob range: start ({start}) > end ({end})"
+        )));
+    }
+    Ok(start..end)
+}
+
+fn parse_u64(value: BigInt, name: &str) -> napi::Result<u64> {
+    let (negative, value, lossless) = value.get_u64();
+    if negative {
+        return Err(napi::Error::from_reason(format!(
+            "{name} cannot be negative"
+        )));
+    }
+    if !lossless {
+        return Err(napi::Error::from_reason(format!(
+            "{name} is too large to fit in u64"
+        )));
+    }
+    Ok(value)
+}
+
+pub fn parse_row_ids(row_ids: Vec<BigInt>) -> napi::Result<Vec<u64>> {
+    row_ids
+        .into_iter()
+        .map(|id| parse_u64(id, "row id"))
+        .collect()
+}
+
+pub fn copy_blob_buffers(array: LargeBinaryArray) -> Vec<Option<Buffer>> {
+    (0..array.len())
+        .map(|i| {
+            if array.is_null(i) {
+                None
+            } else {
+                Some(Buffer::from(array.value(i).to_vec()))
+            }
+        })
+        .collect()
+}

nodejs/src/job.rs

@@ -127,7 +127,7 @@ impl Job {
 }
 
 /// Serialise Arrow batches as a single IPC stream for the TypeScript layer.
-pub(crate) fn batches_to_ipc_buffer(batches: &[RecordBatch]) -> napi::Result<Buffer> {
+fn batches_to_ipc_buffer(batches: &[RecordBatch]) -> napi::Result<Buffer> {
     let Some(first) = batches.first() else {
         return Ok(Buffer::from(Vec::<u8>::new()));
     };

nodejs/src/lib.rs

@@ -10,6 +10,7 @@ use std::collections::HashMap;
 use env_logger::Env;
 use napi_derive::*;
 
+mod blob;
 mod connection;
 mod error;
 mod header;

nodejs/src/table.rs

@@ -15,6 +15,7 @@ use napi::bindgen_prelude::*;
 use napi::threadsafe_function::{ThreadsafeFunction, ThreadsafeFunctionCallMode};
 use napi_derive::napi;
 
+use crate::blob::{BlobFile, copy_blob_buffers, parse_row_ids};
 use crate::error::NapiErrorExt;
 use crate::index::Index;
 use crate::merge::NativeMergeInsertBuilder;
@@ -329,6 +330,44 @@ impl Table {
         ))
     }
 
+    #[napi(catch_unwind)]
+    pub async fn blob_columns(&self) -> napi::Result<Vec<String>> {
+        self.inner_ref()?.blob_columns().await.default_error()
+    }
+
+    #[napi(catch_unwind)]
+    pub async fn fetch_blobs(
+        &self,
+        column: String,
+        row_ids: Vec<BigInt>,
+    ) -> napi::Result<Vec<Option<Buffer>>> {
+        let row_ids = parse_row_ids(row_ids)?;
+        let array = self
+            .inner_ref()?
+            .fetch_blobs(column.as_str(), &row_ids)
+            .await
+            .default_error()?;
+        Ok(copy_blob_buffers(array))
+    }
+
+    #[napi(catch_unwind)]
+    pub async fn fetch_blob_files(
+        &self,
+        column: String,
+        row_ids: Vec<BigInt>,
+    ) -> napi::Result<Vec<Option<BlobFile>>> {
+        let row_ids = parse_row_ids(row_ids)?;
+        let files = self
+            .inner_ref()?
+            .fetch_blob_files(column.as_str(), &row_ids)
+            .await
+            .default_error()?;
+        Ok(files
+            .into_iter()
+            .map(|file| file.map(BlobFile::new))
+            .collect())
+    }
+
     #[napi(catch_unwind)]
     pub fn vector_search(&self, vector: Float32Array) -> napi::Result<VectorQuery> {
         self.query()?.nearest_to(vector)

python/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "lancedb-python"
-version = "0.39.0-beta.2"
+version = "0.39.0-beta.6"
 publish = false
 edition.workspace = true
 description = "Python bindings for LanceDB"

python/python/lancedb/embeddings/gte.py

@@ -21,7 +21,7 @@ class GteEmbeddings(TextEmbeddingFunction):
     An embedding function that uses GTE-LARGE MLX format(for Apple silicon devices only)
     as well as the standard cpu/gpu version from: https://huggingface.co/thenlper/gte-large.
 
-    For Apple users, you will need the mlx package insalled, which can be done with:
+    For Apple users, you will need the mlx package installed, which can be done with:
         pip install mlx
 
     Parameters

python/python/lancedb/embeddings/instructor.py

@@ -60,7 +60,7 @@ class InstructorEmbeddingFunction(TextEmbeddingFunction):
 
     import lancedb
     from lancedb.pydantic import LanceModel, Vector
-    from lancedb.embeddings import get_registry, InstuctorEmbeddingFunction
+    from lancedb.embeddings import get_registry, InstructorEmbeddingFunction
 
     instructor = get_registry().get("instructor").create(
         source_instruction="represent the document for retrieval",

python/python/lancedb/index.py

@@ -751,7 +751,7 @@ class IvfPq:
         This value controls how much the vector is compressed during the
         quantization step.  The more sub vectors there are the less the vector is
         compressed.  The default is the dimension of the vector divided by 16.  If
-        the dimension is not evenly divisible by 16 we use the dimension divded by
+        the dimension is not evenly divisible by 16 we use the dimension divided by
         8.
 
         The above two cases are highly preferred.  Having 8 or 16 values per

python/python/lancedb/query.py

@@ -78,6 +78,10 @@ if TYPE_CHECKING:
 T = TypeVar("T", bound="LanceModel")
 AnalyzePlanDistributedMetrics = Literal["aggregate", "per_worker", "full"]
 
+# Number of rows a hybrid query returns when no limit was set on it. This
+# mirrors the default the Rust query builder applies to its sub-queries.
+DEFAULT_HYBRID_LIMIT = 10
+
 
 @runtime_checkable
 class _LanceScanner(Protocol):
@@ -859,7 +863,7 @@ class Query(pydantic.BaseModel):
         return query
 
     # This tells pydantic to allow custom types (needed for the `vector` query since
-    # pa.Array wouln't be allowed otherwise)
+    # pa.Array wouldn't be allowed otherwise)
     model_config = pydantic.ConfigDict(arbitrary_types_allowed=True)
 
 
@@ -3893,14 +3897,54 @@ class AsyncHybridQuery(AsyncStandardQuery, AsyncVectorQueryBase):
 
         return self
 
+    def _create_child_queries(
+        self,
+    ) -> Tuple["AsyncFTSQuery", "AsyncVectorQuery", int, int]:
+        """Build the sub-queries that make up this hybrid query.
+
+        Execution, `explain_plan` and `analyze_plan` all go through here so that
+        the plans that are reported are the plans that actually run.
+
+        Returns the two sub-queries along with the effective limit and offset of
+        the hybrid query itself.
+        """
+        fts_query = AsyncFTSQuery(self._inner.to_fts_query(), self._table)
+        vec_query = AsyncVectorQuery(self._inner.to_vector_query(), self._table)
+
+        fts_req = fts_query._inner.to_query_request()
+        vec_req = vec_query._inner.to_query_request()
+
+        # Only one of the two sub-queries carries the limit when it was never
+        # set explicitly: nearest_to()/nearest_to_text() build the sibling query
+        # from scratch, and that is where the default gets filled in. Which one
+        # that is depends on the order the hybrid query was built in, so look at
+        # both rather than at a single side.
+        limit = fts_req.limit if fts_req.limit is not None else vec_req.limit
+        if limit is None:
+            limit = DEFAULT_HYBRID_LIMIT
+        offset = fts_req.offset or vec_req.offset or 0
+
+        fts_query.with_row_id()
+        vec_query.with_row_id()
+
+        # offset() pushes the offset down into both sub-queries, which would make
+        # each of them skip its own first `offset` rows. The window has to be
+        # taken out of the combined, reranked results instead, so fetch the
+        # skipped prefix here too and slice it off afterwards.
+        fts_query.limit(limit + offset)
+        vec_query.limit(limit + offset)
+        fts_query.offset(0)
+        vec_query.offset(0)
+
+        return fts_query, vec_query, limit, offset
+
     async def to_batches(
         self,
         *,
         max_batch_length: Optional[int] = None,
         timeout: Optional[timedelta] = None,
     ) -> AsyncRecordBatchReader:
-        fts_query = AsyncFTSQuery(self._inner.to_fts_query(), self._table)
-        vec_query = AsyncVectorQuery(self._inner.to_vector_query(), self._table)
+        fts_query, vec_query, limit, offset = self._create_child_queries()
 
         req = fts_query._inner.to_query_request()
         blob_auto_row_id = False
@@ -3920,9 +3964,6 @@ class AsyncHybridQuery(AsyncStandardQuery, AsyncVectorQueryBase):
         self._blob_auto_row_id = blob_auto_row_id
         self._blob_paths = blob_paths
 
-        fts_query.with_row_id()
-        vec_query.with_row_id()
-
         fts_results, vector_results = await asyncio.gather(
             fts_query.to_arrow(timeout=timeout),
             vec_query.to_arrow(timeout=timeout),
@@ -3934,8 +3975,9 @@ class AsyncHybridQuery(AsyncStandardQuery, AsyncVectorQueryBase):
             norm=self._norm,
             fts_query=fts_query.get_query(),
             reranker=self._reranker,
-            limit=self._inner.get_limit(),
+            limit=limit,
             with_row_ids=True,
+            offset=offset,
         )
         if (
             not self._user_requested_row_id()
@@ -3964,14 +4006,14 @@ class AsyncHybridQuery(AsyncStandardQuery, AsyncVectorQueryBase):
         ...     print(plan)
         >>> asyncio.run(doctest_example()) # doctest: +ELLIPSIS, +NORMALIZE_WHITESPACE
         RRFReranker(K=60)
-            ProjectionExec: expr=[vector@0 as vector, text@3 as text, _distance@2 as _distance]
+            ProjectionExec: expr=[vector@0 as vector, text@3 as text, _distance@2 as _distance, _rowid@1 as _rowid]
               LanceRead: uri=..., projection=[text], source=stream(_rowid)
                 GlobalLimitExec: skip=0, fetch=10
                   FilterExec: _distance@2 IS NOT NULL
                     SortExec: TopK(fetch=10), expr=[_distance@2 ASC NULLS LAST, _rowid@1 ASC NULLS LAST], preserve_partitioning=[false]
                       KNNVectorDistance: metric=l2
                         LanceRead: uri=..., projection=[vector], ...
-            ProjectionExec: expr=[vector@2 as vector, text@3 as text, _score@1 as _score]
+            ProjectionExec: expr=[vector@2 as vector, text@3 as text, _score@1 as _score, _rowid@0 as _rowid]
               LanceRead: uri=..., projection=[vector, text], source=stream(_rowid)
                 GlobalLimitExec: skip=0, fetch=10
                   MatchQuery: column=text, query=[hello]
@@ -3986,8 +4028,9 @@ class AsyncHybridQuery(AsyncStandardQuery, AsyncVectorQueryBase):
         plan : str
         """  # noqa: E501
 
-        vector_plan = await self._inner.to_vector_query().explain_plan(verbose)
-        fts_plan = await self._inner.to_fts_query().explain_plan(verbose)
+        fts_query, vec_query, _, _ = self._create_child_queries()
+        vector_plan = await vec_query.explain_plan(verbose)
+        fts_plan = await fts_query.explain_plan(verbose)
         # Indent sub-plans under the reranker
         indented_vector = "\n".join("  " + line for line in vector_plan.splitlines())
         indented_fts = "\n".join("  " + line for line in fts_plan.splitlines())
@@ -4014,14 +4057,12 @@ class AsyncHybridQuery(AsyncStandardQuery, AsyncVectorQueryBase):
         -------
         plan : str
         """
+        fts_query, vec_query, _, _ = self._create_child_queries()
+
         results = ["Vector Search Query:"]
-        results.append(
-            await self._inner.to_vector_query().analyze_plan(distributed_metrics)
-        )
+        results.append(await vec_query.analyze_plan(distributed_metrics))
         results.append("FTS Search Query:")
-        results.append(
-            await self._inner.to_fts_query().analyze_plan(distributed_metrics)
-        )
+        results.append(await fts_query.analyze_plan(distributed_metrics))
 
         return "\n".join(results)
 

python/python/lancedb/remote/table.py

@@ -720,7 +720,7 @@ class RemoteTable(Table):
         Parameters
         ----------
         query: list/np.ndarray/str/PIL.Image.Image, default None
-            The targetted vector to search for.
+            The targeted vector to search for.
 
             - *default None*.
             Acceptable types are: list, np.ndarray, PIL.Image.Image

python/python/lancedb/rerankers/base.py

@@ -175,7 +175,7 @@ class Reranker(ABC):
             if the results haven't been executed yet or the results in arrow format.
         query : str or None,
             The input query. Some rerankers might not need the query to rerank.
-            In that case, it can be set to None explicitly. This is inteded to
+            In that case, it can be set to None explicitly. This is intended to
             be handled by the reranker implementations.
         deduplicate : bool, optional
             Whether to deduplicate the results based on the `_rowid` column,

python/python/lancedb/table.py

@@ -1619,7 +1619,7 @@ class Table(ABC):
         Parameters
         ----------
         query: list/np.ndarray/str/PIL.Image.Image, default None
-            The targetted vector to search for.
+            The targeted vector to search for.
 
             - *default None*.
             Acceptable types are: list, np.ndarray, PIL.Image.Image
@@ -3841,7 +3841,7 @@ class LanceTable(Table):
         Parameters
         ----------
         query: list/np.ndarray/str/PIL.Image.Image, default None
-            The targetted vector to search for.
+            The targeted vector to search for.
 
             - *default None*.
             Acceptable types are: list, np.ndarray, PIL.Image.Image
@@ -5638,7 +5638,7 @@ class AsyncTable:
         if fill_value is None:
             fill_value = 0.0
 
-        # _santitize_data is an old code path, but we will use it until the
+        # _sanitize_data is an old code path, but we will use it until the
         # new code path is ready.
         if mode == "overwrite":
             # For overwrite, apply the same preprocessing as create_table
@@ -5814,7 +5814,7 @@ class AsyncTable:
         Parameters
         ----------
         query: list/np.ndarray/str/PIL.Image.Image, default None
-            The targetted vector to search for.
+            The targeted vector to search for.
 
             - *default None*.
             Acceptable types are: list, np.ndarray, PIL.Image.Image

python/python/tests/test_blob.py

@@ -297,7 +297,10 @@ def test_blob_v2_projection_sources_use_typed_column_name():
 
 
 def _legacy_v1_table(name):
-    db = lancedb.connect("memory:///")
+    # Legacy v1 blob columns are only writable at file version <= 2.1.
+    db = lancedb.connect(
+        "memory:///", storage_options={"new_table_data_storage_version": "2.1"}
+    )
     schema = pa.schema(
         [
             pa.field("id", pa.int64()),

python/python/tests/test_embeddings.py

@@ -327,8 +327,8 @@ def test_embedding_function_with_pandas(tmp_path):
         ) -> List[np.array]:
             return [np.random.randn(self.ndims()).tolist() for _ in range(len(texts))]
 
-    registery = get_registry()
-    func = registery.get("mock-embedding").create()
+    registry = get_registry()
+    func = registry.get("mock-embedding").create()
 
     class TestSchema(LanceModel):
         text: str = func.SourceField()
@@ -394,9 +394,9 @@ def test_multiple_embeddings_for_pandas(tmp_path):
         ) -> List[np.array]:
             return [np.random.randn(self.ndims()).tolist() for _ in range(len(texts))]
 
-    registery = get_registry()
-    func1 = registery.get("mock-embedding").create()
-    func2 = registery.get("mock-embedding2").create()
+    registry = get_registry()
+    func1 = registry.get("mock-embedding").create()
+    func2 = registry.get("mock-embedding2").create()
 
     class TestSchema(LanceModel):
         text: str = func1.SourceField()

python/python/tests/test_fts.py

@@ -1011,8 +1011,13 @@ def test_fts_ngram(mem_db: DBConnection):
     assert set(r["text"] for r in results) == {"lance database", "lance is cool"}
 
     results = (
-        table.search("nce", query_type="fts").limit(10).to_list()
-    )  # spellchecker:disable-line
+        table.search(
+            "nce",  # spellchecker:disable-line
+            query_type="fts",
+        )
+        .limit(10)
+        .to_list()
+    )
     assert len(results) == 2
     assert set(r["text"] for r in results) == {"lance database", "lance is cool"}
 
@@ -1034,8 +1039,13 @@ def test_fts_ngram(mem_db: DBConnection):
     assert set(r["text"] for r in results) == {"lance database", "lance is cool"}
 
     results = (
-        table.search("nce", query_type="fts").limit(10).to_list()
-    )  # spellchecker:disable-line
+        table.search(
+            "nce",  # spellchecker:disable-line
+            query_type="fts",
+        )
+        .limit(10)
+        .to_list()
+    )
     assert len(results) == 0
 
     results = table.search("la", query_type="fts").limit(10).to_list()

python/python/tests/test_hybrid_query.py

@@ -203,6 +203,93 @@ async def test_async_hybrid_query_default_limit(table: AsyncTable):
     assert texts.count("a") == 1
 
 
+@pytest.mark.asyncio
+async def test_async_hybrid_query_offset(table: AsyncTable):
+    # The offset window of a hybrid query must be a suffix of the same query
+    # run without an offset. Skipping the first rows of each sub-query instead
+    # of the first rows of the fused result silently changes which rows land in
+    # the window.
+    full = await (
+        table.query()
+        .nearest_to([0.0, 0.4])
+        .nearest_to_text("dog")
+        .limit(4)
+        .with_row_id()
+        .to_arrow()
+    )
+    assert len(full) == 4
+
+    second_page = await (
+        table.query()
+        .nearest_to([0.0, 0.4])
+        .nearest_to_text("dog")
+        .offset(2)
+        .limit(2)
+        .with_row_id()
+        .to_arrow()
+    )
+    assert second_page["_rowid"].to_pylist() == full["_rowid"].to_pylist()[2:]
+
+    first_page = await (
+        table.query()
+        .nearest_to([0.0, 0.4])
+        .nearest_to_text("dog")
+        .limit(2)
+        .with_row_id()
+        .to_arrow()
+    )
+    # Paging through the result must visit every row exactly once: no row
+    # repeated from the previous page and none dropped between the two.
+    paged = first_page["_rowid"].to_pylist() + second_page["_rowid"].to_pylist()
+    assert sorted(paged) == sorted(full["_rowid"].to_pylist())
+
+
+@pytest.mark.asyncio
+async def test_async_hybrid_query_fts_first_default_limit(table: AsyncTable):
+    # nearest_to() and nearest_to_text() build their new sibling sub-query from
+    # scratch, and that is the sub-query the default limit ends up on. So the
+    # side that carries the limit depends on the order the hybrid query was
+    # built in, and looking at only one side loses the limit for half the ways
+    # a hybrid query can be written. Without a limit the combined results are
+    # not truncated at all and the whole union of both candidate lists is
+    # returned.
+    await table.add([{"text": "dog", "vector": [50.0 + i, 50.0]} for i in range(10)])
+
+    result = await (
+        table.query().nearest_to_text("dog").nearest_to([0.1, 0.1]).to_arrow()
+    )
+    assert len(result) == 10
+
+    offset_result = await (
+        table.query().nearest_to_text("dog").nearest_to([0.1, 0.1]).offset(2).to_arrow()
+    )
+    assert len(offset_result) == 10
+
+
+@pytest.mark.asyncio
+async def test_async_hybrid_query_explain_plan_matches_execution(table: AsyncTable):
+    # Paging rewrites the sub-queries: each one fetches limit + offset rows with
+    # no offset of its own, and the window is sliced out after fusion. The plans
+    # have to be built from those rewritten sub-queries, otherwise explain_plan
+    # and analyze_plan describe a query that is never run.
+    query = (
+        table.query().nearest_to([0.0, 0.4]).nearest_to_text("dog").offset(2).limit(2)
+    )
+    await query.to_arrow()
+
+    plan = await query.explain_plan()
+    assert [
+        line.strip() for line in plan.splitlines() if "GlobalLimitExec" in line
+    ] == [
+        "GlobalLimitExec: skip=0, fetch=4",
+        "GlobalLimitExec: skip=0, fetch=4",
+    ]
+
+    analyzed = await query.analyze_plan()
+    assert analyzed.count("skip=0, fetch=4") == 2
+    assert "skip=2" not in analyzed
+
+
 def test_hybrid_query_offset(sync_table: Table):
     # The offset window of a hybrid query must be a suffix of the same query
     # run without an offset -- it must not be silently ignored.

python/python/tests/test_namespace.py

@@ -193,7 +193,13 @@ class TestNamespaceConnection:
             ),
         )
 
-        table = db.create_table("blob_table", data, namespace_path=["test_ns"])
+        # Legacy v1 blob columns are only writable at file version <= 2.1.
+        table = db.create_table(
+            "blob_table",
+            data,
+            namespace_path=["test_ns"],
+            storage_options={"new_table_data_storage_version": "2.1"},
+        )
         df = table.to_pandas(blob_mode="lazy").sort_values("id")
 
         blob = df["blob"].iloc[0]

python/python/tests/test_query.py

@@ -40,6 +40,10 @@ from utils import exception_output
 from importlib.util import find_spec
 
 
+# Legacy v1 blob columns are only writable at file version <= 2.1.
+LEGACY_BLOB_STORAGE_OPTIONS = {"new_table_data_storage_version": "2.1"}
+
+
 def _blob_query_data():
     return pa.table(
         {
@@ -119,13 +123,17 @@ def _assert_blob_bytes_projection(df):
 
 def _blob_query_table(db, name, blob_schema):
     if blob_schema == "v1":
-        return db.create_table(name, _blob_query_data())
+        return db.create_table(
+            name, _blob_query_data(), storage_options=LEGACY_BLOB_STORAGE_OPTIONS
+        )
     return _create_blob_v2_query_table(db, name)
 
 
 async def _blob_query_table_async(db, name, blob_schema):
     if blob_schema == "v1":
-        return await db.create_table(name, _blob_query_data())
+        return await db.create_table(
+            name, _blob_query_data(), storage_options=LEGACY_BLOB_STORAGE_OPTIONS
+        )
     return await _create_blob_v2_query_table_async(db, name)
 
 
@@ -275,7 +283,9 @@ async def test_query_to_pandas_kwargs(table, table_async):
 def test_plain_scan_query_to_pandas_blob_modes(tmp_db, blob_mode):
     pytest.importorskip("lance")
     table = tmp_db.create_table(
-        f"test_query_to_pandas_blob_{blob_mode}", _blob_query_data()
+        f"test_query_to_pandas_blob_{blob_mode}",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
     )
 
     df = (
@@ -322,7 +332,9 @@ def test_plain_scan_query_to_pandas_blob_mode_does_not_collect_arrow(
 ):
     pytest.importorskip("lance")
     table = tmp_db.create_table(
-        "test_query_to_pandas_blob_no_arrow_collect", _blob_query_data()
+        "test_query_to_pandas_blob_no_arrow_collect",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
     )
     query = table.search().where("id = 1").select(["id", "blob"])
 
@@ -347,7 +359,9 @@ def test_plain_scan_query_to_pandas_blob_descriptions_flatten_uses_scanner(
 ):
     pytest.importorskip("lance")
     table = tmp_db.create_table(
-        "test_query_to_pandas_blob_desc_flatten", _blob_query_data()
+        "test_query_to_pandas_blob_desc_flatten",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
     )
     query = table.search().where("id = 1").select(["id", "blob"])
 
@@ -365,7 +379,11 @@ def test_plain_scan_query_to_pandas_blob_descriptions_flatten_uses_scanner(
 def test_plain_scan_query_to_pandas_scanner_state(tmp_db):
     pytest.importorskip("lance")
     data = _blob_query_data()
-    table = tmp_db.create_table("test_query_to_pandas_scanner_state", data.slice(0, 2))
+    table = tmp_db.create_table(
+        "test_query_to_pandas_scanner_state",
+        data.slice(0, 2),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
+    )
     table.add(data.slice(2, 2))
 
     fragments = table.to_lance().get_fragments()
@@ -400,7 +418,9 @@ def test_plain_scan_query_to_pandas_scanner_state(tmp_db):
 async def test_async_plain_scan_query_to_pandas_blob_projection(tmp_db_async):
     pytest.importorskip("lance")
     table = await tmp_db_async.create_table(
-        "test_async_query_to_pandas_blob_projection", _blob_query_data()
+        "test_async_query_to_pandas_blob_projection",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
     )
 
     lazy_df = await (
@@ -452,7 +472,9 @@ async def test_async_plain_scan_query_to_pandas_blob_mode_does_not_collect_arrow
 ):
     pytest.importorskip("lance")
     table = await tmp_db_async.create_table(
-        "test_async_query_to_pandas_blob_no_arrow_collect", _blob_query_data()
+        "test_async_query_to_pandas_blob_no_arrow_collect",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
     )
     query = table.query().where("id = 1").select(["id", "blob"])
 
@@ -474,7 +496,11 @@ async def test_async_plain_scan_query_to_pandas_blob_mode_does_not_collect_arrow
 
 def test_vector_query_to_pandas_blob_mode_requires_native_path(tmp_db):
     pytest.importorskip("lance")
-    table = tmp_db.create_table("test_vector_query_blob_mode", _blob_query_data())
+    table = tmp_db.create_table(
+        "test_vector_query_blob_mode",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
+    )
 
     with pytest.raises(RuntimeError, match="Lance native pandas conversion"):
         table.search([1.0, 0.0]).select(["blob", "vector"]).limit(1).to_pandas(
@@ -485,7 +511,9 @@ def test_vector_query_to_pandas_blob_mode_requires_native_path(tmp_db):
 def test_vector_query_to_pandas_blob_descriptions_requires_plain_scan(tmp_db):
     pytest.importorskip("lance")
     table = tmp_db.create_table(
-        "test_vector_query_blob_descriptions", _blob_query_data()
+        "test_vector_query_blob_descriptions",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
     )
 
     with pytest.raises(RuntimeError, match="plain scan query"):

python/python/tests/test_rerankers.py

@@ -81,7 +81,7 @@ def get_test_table(tmp_path):
         "but his son was mortal",
         "there hasn't been a good battlefield game since 2142",
         "I wish they would make another one",
-        "campains are not as good as they used to be",
+        "campaigns are not as good as they used to be",
         "Multiplayer and open world games have destroyed the single player experience",
         "Maybe the future is console games",
         "I don't know",

python/python/tests/test_table.py

@@ -64,15 +64,23 @@ async def _blob_v2_table_async(db: AsyncConnection, name: str):
     return table
 
 
+# Legacy v1 blob columns are only writable at file version <= 2.1.
+LEGACY_BLOB_STORAGE_OPTIONS = {"new_table_data_storage_version": "2.1"}
+
+
 def _blob_table(db: DBConnection, name: str, blob_schema: str):
     if blob_schema == "v1":
-        return db.create_table(name, data=_blob_test_data())
+        return db.create_table(
+            name, data=_blob_test_data(), storage_options=LEGACY_BLOB_STORAGE_OPTIONS
+        )
     return _blob_v2_table(db, name)
 
 
 async def _blob_table_async(db: AsyncConnection, name: str, blob_schema: str):
     if blob_schema == "v1":
-        return await db.create_table(name, data=_blob_test_data())
+        return await db.create_table(
+            name, data=_blob_test_data(), storage_options=LEGACY_BLOB_STORAGE_OPTIONS
+        )
     return await _blob_v2_table_async(db, name)
 
 
@@ -147,7 +155,11 @@ def test_table_to_pandas_invalid_blob_mode_non_blob_table(tmp_db: DBConnection):
 @pytest.mark.parametrize("blob_mode", ["lazy", "bytes", "descriptions"])
 def test_table_to_pandas_blob_modes(tmp_db: DBConnection, blob_mode):
     pytest.importorskip("lance")
-    table = tmp_db.create_table(f"test_to_pandas_blob_{blob_mode}", _blob_test_data())
+    table = tmp_db.create_table(
+        f"test_to_pandas_blob_{blob_mode}",
+        _blob_test_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
+    )
 
     df = table.to_pandas(blob_mode=blob_mode)
 
@@ -3342,7 +3354,7 @@ def test_empty_query(mem_db: DBConnection):
     # None is the same as default
     df = table.search().select(["id"]).limit(None).to_arrow()
     assert df.num_rows == 100
-    # invalid limist is the same as None, wihch is the same as default
+    # invalid limist is the same as None, which is the same as default
     df = table.search().select(["id"]).limit(-1).to_arrow()
     assert df.num_rows == 100
     # valid limit should work
@@ -3959,7 +3971,7 @@ def test_stats(mem_db: DBConnection):
     print(f"{stats=}")
     assert stats == {
         # Full on-disk size of the data file, footer and metadata included.
-        "total_bytes": 633,
+        "total_bytes": 637,
         "num_rows": 2,
         "num_indices": 0,
         "fragment_stats": {

python/src/query.rs

@@ -334,7 +334,7 @@ pub struct PyQueryRequest {
     pub column: Option<String>,
     pub query_vector: Option<PyQueryVectors>,
     pub minimum_nprobes: Option<usize>,
-    // None means user did not set it and default shoud be used (currenty 20)
+    // None means user did not set it and default should be used (currently 20)
     // Some(0) means user set it to None and there is no limit
     pub maximum_nprobes: Option<usize>,
     pub lower_bound: Option<f32>,

rust/lancedb/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "lancedb"
-version = "0.39.0-beta.2"
+version = "0.39.0-beta.6"
 edition.workspace = true
 description = "LanceDB: A serverless, low-latency vector database for AI applications"
 license.workspace = true

rust/lancedb/src/arrow.rs

@@ -163,7 +163,7 @@ pub struct PolarsDataFrameRecordBatchReader {
 impl PolarsDataFrameRecordBatchReader {
     /// Creates a new `PolarsDataFrameRecordBatchReader` from a given Polars DataFrame.
     /// If the input dataframe does not have aligned chunks, this function undergoes
-    /// the costly operation of reallocating each series as a single contigous chunk.
+    /// the costly operation of reallocating each series as a single contiguous chunk.
     pub fn new(mut df: DataFrame) -> Result<Self> {
         df.align_chunks();
         let arrow_schema =

rust/lancedb/src/blob.rs

@@ -532,10 +532,11 @@ mod tests {
     fn storage_version_bumps_to_v2_2() {
         let mut params = WriteParams::default();
         ensure_blob_storage_version(&blob_schema(), &mut params);
-        assert_eq!(
-            params.data_storage_version.unwrap().resolve(),
-            ConcreteFileVersion::V2_2
-        );
+        let resolved = params
+            .data_storage_version
+            .unwrap_or(LanceFileVersion::Stable)
+            .resolve();
+        assert_eq!(resolved, ConcreteFileVersion::V2_2);
         assert!(!params.enable_stable_row_ids);
     }
 
@@ -547,10 +548,11 @@ mod tests {
         };
         ensure_blob_storage_version(&blob_schema(), &mut params);
         assert!(params.enable_stable_row_ids);
-        assert_eq!(
-            params.data_storage_version.unwrap().resolve(),
-            ConcreteFileVersion::V2_2
-        );
+        let resolved = params
+            .data_storage_version
+            .unwrap_or(LanceFileVersion::Stable)
+            .resolve();
+        assert_eq!(resolved, ConcreteFileVersion::V2_2);
     }
 
     #[test]

rust/lancedb/src/connection.rs

@@ -827,7 +827,7 @@ impl Connection {
 pub struct ConnectRequest {
     /// Database URI
     ///
-    /// ### Accpeted URI formats
+    /// ### Accepted URI formats
     ///
     /// - `/path/to/database` - local database on file system.
     /// - `s3://bucket/path/to/database` or `gs://bucket/path/to/database` - database on cloud object store

rust/lancedb/src/database/listing.rs

@@ -512,7 +512,7 @@ impl ListingDatabase {
                 // iter thru the query params and extract the commit store param
                 let mut engine = None;
                 let mut mirrored_store = None;
-                let mut filtered_querys = vec![];
+                let mut filtered_queries = vec![];
 
                 // WARNING: specifying engine is NOT a publicly supported feature in lancedb yet
                 // THE API WILL CHANGE
@@ -528,13 +528,13 @@ impl ListingDatabase {
                         mirrored_store = Some(value.to_string());
                     } else {
                         // to owned so we can modify the url
-                        filtered_querys.push((key.to_string(), value.to_string()));
+                        filtered_queries.push((key.to_string(), value.to_string()));
                     }
                 }
 
                 // Filter out the commit store query param -- it's a lancedb param
                 url.query_pairs_mut().clear();
-                url.query_pairs_mut().extend_pairs(filtered_querys);
+                url.query_pairs_mut().extend_pairs(filtered_queries);
                 // Take a copy of the query string so we can propagate it to lance.
                 // `query_pairs_mut()` leaves the URL with `Some("")` even when no
                 // pairs survive (or none existed in the first place), so an empty
@@ -896,11 +896,11 @@ impl Database for ListingDatabase {
     }
 
     async fn read_consistency(&self) -> Result<ReadConsistency> {
-        if let Some(read_consistency_inverval) = self.read_consistency_interval {
-            if read_consistency_inverval.is_zero() {
+        if let Some(interval) = self.read_consistency_interval {
+            if interval.is_zero() {
                 Ok(ReadConsistency::Strong)
             } else {
-                Ok(ReadConsistency::Eventual(read_consistency_inverval))
+                Ok(ReadConsistency::Eventual(interval))
             }
         } else {
             Ok(ReadConsistency::Manual)
@@ -3043,15 +3043,15 @@ mod tests {
     /// across platforms — see the `file://` test below).
     fn capture_query_like_connect(input_uri: &str) -> Option<String> {
         let mut url = url::Url::parse(input_uri).unwrap();
-        let mut filtered_querys = Vec::new();
+        let mut filtered_queries = Vec::new();
         for (key, value) in url.query_pairs() {
             if key == ENGINE || key == MIRRORED_STORE {
                 continue;
             }
-            filtered_querys.push((key.to_string(), value.to_string()));
+            filtered_queries.push((key.to_string(), value.to_string()));
         }
         url.query_pairs_mut().clear();
-        url.query_pairs_mut().extend_pairs(filtered_querys);
+        url.query_pairs_mut().extend_pairs(filtered_queries);
         url.query().filter(|q| !q.is_empty()).map(|s| s.to_string())
     }
 

rust/lancedb/src/database/namespace.rs

@@ -3,6 +3,7 @@
 
 //! Namespace-based database implementation that delegates table management to lance-namespace
 
+use lance_datafusion::utils::StreamingWriteSource;
 use std::collections::{HashMap, HashSet};
 use std::sync::{Arc, Mutex};
 
@@ -250,11 +251,11 @@ impl Database for LanceNamespaceDatabase {
     }
 
     async fn read_consistency(&self) -> Result<ReadConsistency> {
-        if let Some(read_consistency_inverval) = self.read_consistency_interval {
-            if read_consistency_inverval.is_zero() {
+        if let Some(interval) = self.read_consistency_interval {
+            if interval.is_zero() {
                 Ok(ReadConsistency::Strong)
             } else {
-                Ok(ReadConsistency::Eventual(read_consistency_inverval))
+                Ok(ReadConsistency::Eventual(interval))
             }
         } else {
             Ok(ReadConsistency::Manual)
@@ -304,6 +305,10 @@ impl Database for LanceNamespaceDatabase {
     }
 
     async fn create_table(&self, request: DbCreateTableRequest) -> Result<Arc<dyn BaseTable>> {
+        // Refuse a bad declaration before the namespace records a table.
+        crate::table::computed_columns::ensure_declarations_are_planned(
+            &request.data.arrow_schema(),
+        )?;
         let mut table_id = request.namespace_path.clone();
         table_id.push(request.name.clone());
         let mut existing_table = None;

rust/lancedb/src/index/vector.rs

@@ -125,7 +125,7 @@ macro_rules! impl_pq_params_setter {
         /// This value controls how much the vector is compressed during the quantization step.
         /// The more sub vectors there are the less the vector is compressed.  The default is
         /// the dimension of the vector divided by 16.  If the dimension is not evenly divisible
-        /// by 16 we use the dimension divded by 8.
+        /// by 16 we use the dimension divided by 8.
         ///
         /// The above two cases are highly preferred.  Having 8 or 16 values per subvector allows
         /// us to use efficient SIMD instructions.

rust/lancedb/src/materialized_view/refresh.rs

@@ -24,8 +24,8 @@ use std::time::{SystemTime, UNIX_EPOCH};
 
 use arrow_array::cast::AsArray;
 use arrow_array::types::UInt64Type;
-use arrow_array::{RecordBatch, UInt64Array};
-use arrow_schema::{Schema as ArrowSchema, SchemaRef};
+use arrow_array::{RecordBatch, UInt64Array, new_null_array};
+use arrow_schema::{FieldRef, Schema as ArrowSchema, SchemaRef};
 use datafusion::common::ScalarValue;
 use datafusion::error::DataFusionError;
 use datafusion::physical_plan::SendableRecordBatchStream;
@@ -34,7 +34,7 @@ use datafusion::prelude::{col, lit};
 use futures::{StreamExt, TryStreamExt};
 use lance::Dataset;
 use lance::dataset::mem_wal::DatasetMemWalExt;
-use lance::dataset::transaction::{Operation, Transaction};
+use lance::dataset::transaction::{Operation, Transaction, UpdateMode};
 use lance::dataset::write::delete::DeleteBuilder;
 use lance::dataset::write::merge_insert::inserted_rows::{
     KeyExistenceFilter, KeyExistenceFilterBuilder, KeyValue,
@@ -51,6 +51,9 @@ use super::{
     definition_to_metadata,
 };
 use crate::database::OpenTableRequest;
+use crate::table::computed_columns::{
+    computed_column_from_field, computed_columns, ensure_declarations_are_planned,
+};
 use crate::table::{NativeTable, NativeTableExt, Table};
 use crate::{Error, Result};
 
@@ -167,30 +170,52 @@ pub(crate) async fn execute_refresh(
         .map(|p| (p.output.clone(), p.expression.clone()))
         .collect();
     validate_inputs(&source_ds, definition)?;
-    let (replanned, mut planned_fields, _renames) = super::plan(
+    let (replanned, planned_fields, _renames) = super::plan(
         source_schema,
         &definition.source_table,
         &definition.source_namespace,
-        &projections,
+        Some(&projections),
         definition.filter.as_deref(),
         definition.limit,
     )?;
+    let mut planned_fields = planned_fields;
     planned_fields.push(arrow_schema::Field::new(
         SOURCE_ROW_ID_COLUMN,
         arrow_schema::DataType::UInt64,
         false,
     ));
+    // A computed column is not planned from the source: refresh writes it
+    // NULL and its declaration's owner fills it. Its declaration must still
+    // be complete, and it must be able to hold NULL.
     let physical = ArrowSchema::from(view_ds.schema());
-    let planned_shape: Vec<_> = planned_fields
-        .iter()
-        .map(|f| (f.name().clone(), f.data_type().clone(), f.is_nullable()))
-        .collect();
-    let physical_shape: Vec<_> = physical
+    let mut computed = computed_columns(&physical).into_iter().map(|c| c.name);
+    if let Some(name) = computed.by_ref().find(|name| {
+        physical
+            .field_with_name(name)
+            .is_ok_and(|f| !f.is_nullable())
+    }) {
+        return Err(Error::Schema {
+            message: format!(
+                "computed column '{name}' of view '{}' cannot hold NULL; recreate the view",
+                view.name()
+            ),
+        });
+    }
+    ensure_declarations_are_planned(&physical)?;
+    let physical_planned: Vec<&FieldRef> = physical
         .fields()
         .iter()
-        .map(|f| (f.name().clone(), f.data_type().clone(), f.is_nullable()))
+        .filter(|f| computed_column_from_field(f).is_none())
         .collect();
-    if planned_shape != physical_shape {
+    // A projected column that became nullable at the source still fits the
+    // view's nullable field; the reverse would not.
+    let matches = planned_fields.len() == physical_planned.len()
+        && planned_fields.iter().zip(&physical_planned).all(|(e, p)| {
+            e.name() == p.name()
+                && e.data_type() == p.data_type()
+                && (p.is_nullable() || !e.is_nullable())
+        });
+    if !matches {
         return Err(Error::Schema {
             message: format!(
                 "the stored definition of view '{}' does not produce this \
@@ -229,11 +254,18 @@ pub(crate) async fn execute_refresh(
         .get(SOURCE_VERSION_TS_META_KEY)
         .and_then(|raw| raw.parse().ok());
     // The watermark speaks only for the view state its refresh left behind;
-    // any other commit on the view since then is drift.
-    let view_intact = metadata
+    // any other commit on the view since then is drift, except a fill of its
+    // computed columns, which rewrites nothing refresh certifies.
+    let recorded_view_version = metadata
         .get(VIEW_VERSION_META_KEY)
-        .and_then(|raw| raw.parse::<u64>().ok())
-        == Some(view_ds.version().version);
+        .and_then(|raw| raw.parse::<u64>().ok());
+    let view_intact = match recorded_view_version {
+        Some(recorded) if recorded == view_ds.version().version => true,
+        Some(recorded) if recorded < view_ds.version().version => {
+            only_computed_rewrites_since(&view_ds, recorded).await?
+        }
+        _ => false,
+    };
 
     if !full && watermark == Some(source_version) && view_intact && recorded_ts == Some(source_ts) {
         return Ok(RefreshMaterializedViewResult {
@@ -1090,6 +1122,69 @@ struct RowScope {
     limit: Option<u64>,
 }
 
+/// Whether every commit on the view after `recorded` is a fill of its
+/// computed columns: a column rewrite or data replacement touching only
+/// those fields and neither adding nor removing rows. A version whose
+/// transaction cannot be read is not proven, so it counts as drift.
+async fn only_computed_rewrites_since(view_ds: &Dataset, recorded: u64) -> Result<bool> {
+    // A fill may write any field under a computed column, so the whole
+    // subtree counts, not only the root.
+    let physical = ArrowSchema::from(view_ds.schema());
+    fn subtree(field: &lance_core::datatypes::Field, ids: &mut Vec<u32>) {
+        ids.push(field.id as u32);
+        for child in &field.children {
+            subtree(child, ids);
+        }
+    }
+    let mut computed_fields = Vec::new();
+    for column in computed_columns(&physical) {
+        if let Some(field) = view_ds.schema().field(&column.name) {
+            subtree(field, &mut computed_fields);
+        }
+    }
+    if computed_fields.is_empty() {
+        return Ok(false);
+    }
+    for version in recorded + 1..=view_ds.version().version {
+        let Some(transaction) = view_ds.read_transaction_by_version(version).await? else {
+            return Ok(false);
+        };
+        let fill = match &transaction.operation {
+            Operation::Update {
+                removed_fragment_ids,
+                new_fragments,
+                fields_modified,
+                update_mode: Some(UpdateMode::RewriteColumns),
+                ..
+            } => {
+                removed_fragment_ids.is_empty()
+                    && new_fragments.is_empty()
+                    && !fields_modified.is_empty()
+                    && fields_modified
+                        .iter()
+                        .all(|field| computed_fields.contains(field))
+            }
+            // What `refresh_column` commits for a SQL declaration.
+            Operation::DataReplacement { replacements } => {
+                !replacements.is_empty()
+                    && replacements.iter().all(|group| {
+                        !group.1.fields.is_empty()
+                            && group
+                                .1
+                                .fields
+                                .iter()
+                                .all(|field| computed_fields.contains(&(*field as u32)))
+                    })
+            }
+            _ => false,
+        };
+        if !fill {
+            return Ok(false);
+        }
+    }
+    Ok(true)
+}
+
 async fn compute_stream(
     source: &Dataset,
     definition: &MaterializedViewDefinition,
@@ -1158,6 +1253,10 @@ async fn compute_stream(
         let batch = batch.map_err(|e| DataFusionError::External(Box::new(e)))?;
         let mut columns = Vec::with_capacity(out_schema.fields().len());
         for field in out_schema.fields() {
+            if computed_column_from_field(field).is_some() {
+                columns.push(new_null_array(field.data_type(), batch.num_rows()));
+                continue;
+            }
             let name = if field.name() == SOURCE_ROW_ID_COLUMN {
                 ROW_ID
             } else {
@@ -2768,7 +2867,7 @@ mod tests {
         let (conn, source) = db_with_source(vec![1]).await;
         let prepared = crate::materialized_view::prepare_declaration(
             &source,
-            &[("x".into(), "x".into()), ("twice".into(), "x * 2".into())],
+            Some(&[("x".into(), "x".into()), ("twice".into(), "x * 2".into())]),
             None,
             None,
         )
@@ -3132,4 +3231,424 @@ mod tests {
         let err = view.refresh().execute().await.unwrap_err();
         assert!(err.to_string().contains("source table 'src'"), "{err}");
     }
+
+    /// A view with a computed column, declared over `people` and refreshed.
+    async fn refreshed_computed_view(conn: &Connection) -> MaterializedView {
+        use crate::materialized_view::tests::{computed_field, people, test_binding};
+        let source = people(conn).await;
+        let view = crate::materialized_view::prepare_declaration(
+            &source,
+            Some(&[
+                ("id".to_string(), "id".to_string()),
+                ("name".to_string(), "name".to_string()),
+            ]),
+            None,
+            None,
+        )
+        .await
+        .unwrap()
+        .with_computed_columns(
+            vec![(2, computed_field("emb", "fb_1", "name"))],
+            &[test_binding("fb_1", "name", "emb")],
+        )
+        .unwrap()
+        .create("v")
+        .await
+        .unwrap();
+        let result = view.refresh().execute().await.unwrap();
+        assert_eq!(result.mode, RefreshMode::Rebuild);
+        view
+    }
+
+    async fn unfilled(view: &MaterializedView) -> usize {
+        view.table()
+            .count_rows(Some("emb IS NULL".to_string()))
+            .await
+            .unwrap()
+    }
+
+    async fn append_people(conn: &Connection, ids: Vec<i32>, names: Vec<&str>) {
+        let batch = record_batch!(("id", Int32, ids), ("name", Utf8, names)).unwrap();
+        conn.open_table("people")
+            .execute()
+            .await
+            .unwrap()
+            .add(batch)
+            .execute()
+            .await
+            .unwrap();
+    }
+
+    /// Commit the fill job's shape on the view: a column rewrite of
+    /// `fields`, touching no rows. The data is left as it is; what matters
+    /// here is how the next refresh classifies the commit.
+    async fn commit_column_rewrite(view: &MaterializedView, fields: &[&str]) {
+        let native = view.table().as_native().unwrap();
+        native.dataset.reload().await.unwrap();
+        let dataset = native.dataset.get().await.unwrap().as_ref().clone();
+        let fields_modified = fields
+            .iter()
+            .map(|name| dataset.schema().field(name).unwrap().id as u32)
+            .collect();
+        let updated_fragments = dataset
+            .get_fragments()
+            .iter()
+            .map(|fragment| fragment.metadata().clone())
+            .collect();
+        let operation = Operation::Update {
+            removed_fragment_ids: Vec::new(),
+            updated_fragments,
+            new_fragments: Vec::new(),
+            fields_modified,
+            compacted_sstables: Vec::new(),
+            fields_for_preserving_frag_bitmap: Vec::new(),
+            update_mode: Some(UpdateMode::RewriteColumns),
+            inserted_rows_filter: None,
+            updated_fragment_offsets: None,
+        };
+        let read_version = dataset.version().version;
+        CommitBuilder::new(WriteDestination::Dataset(Arc::new(dataset)))
+            .execute(Transaction::new(read_version, operation, None))
+            .await
+            .unwrap();
+    }
+
+    /// Refresh never computes a computed column: every row it writes, on a
+    /// rebuild, an append and a rewrite, carries NULL there, and the
+    /// declaration survives all three.
+    #[tokio::test]
+    async fn test_computed_columns_are_written_null_and_kept() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+        assert_eq!(unfilled(&view).await, 3);
+
+        append_people(&conn, vec![4, 5], vec!["d", "e"]).await;
+        let result = view.refresh().execute().await.unwrap();
+        assert_eq!(result.mode, RefreshMode::Incremental);
+        assert_eq!(unfilled(&view).await, 5);
+
+        conn.open_table("people")
+            .execute()
+            .await
+            .unwrap()
+            .update()
+            .column("name", "'z'")
+            .only_if("id = 1")
+            .execute()
+            .await
+            .unwrap();
+        view.refresh().execute().await.unwrap();
+        assert_eq!(unfilled(&view).await, 5);
+        assert_eq!(read(view.table(), "id").await, vec![1, 2, 3, 4, 5]);
+
+        let schema = view.table().schema().await.unwrap();
+        assert!(
+            crate::table::computed_columns::function_bindings(&schema)
+                .unwrap()
+                .iter()
+                .any(|b| b.binding_id() == "fb_1"),
+            "the binding envelope was lost"
+        );
+        assert!(
+            computed_column_from_field(schema.field_with_name("emb").unwrap()).is_some(),
+            "the declaration was lost"
+        );
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+    }
+
+    /// The fill job's commit rewrites only computed columns. It is the one
+    /// commit on a view that is not drift: the next refresh carries on from
+    /// its watermark instead of rebuilding, which would null what the fill
+    /// just wrote.
+    #[tokio::test]
+    async fn test_a_computed_column_fill_is_not_drift() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+
+        commit_column_rewrite(&view, &["emb"]).await;
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+
+        commit_column_rewrite(&view, &["emb"]).await;
+        append_people(&conn, vec![4], vec!["d"]).await;
+        let result = view.refresh().execute().await.unwrap();
+        assert_eq!(result.mode, RefreshMode::Incremental);
+        assert_eq!(result.rows_written, 1);
+        assert_eq!(read(view.table(), "id").await, vec![1, 2, 3, 4]);
+    }
+
+    /// A column rewrite that reaches a projected column is drift like any
+    /// other write: refresh certifies those columns and must recompute them.
+    #[tokio::test]
+    async fn test_a_rewrite_of_a_projected_column_is_drift() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+
+        commit_column_rewrite(&view, &["emb", "name"]).await;
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Rebuild
+        );
+    }
+
+    /// The declaration contract is checked before any refresh mutation: a
+    /// missing binding envelope and a column that lost its declaration both
+    /// fail closed.
+    #[tokio::test]
+    async fn test_a_broken_declaration_is_refused_before_refresh() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+        let native = view.table().as_native().unwrap();
+        let mut dataset = native.dataset.get().await.unwrap().as_ref().clone();
+        dataset
+            .update_schema_metadata(vec![(
+                crate::table::computed_columns::FUNCTION_BINDINGS_META_KEY.to_string(),
+                None,
+            )])
+            .await
+            .unwrap();
+        let err = view.refresh().execute().await.unwrap_err().to_string();
+        assert!(err.contains("references missing binding 'fb_1'"), "{err}");
+
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+        let native = view.table().as_native().unwrap();
+        let mut dataset = native.dataset.get().await.unwrap().as_ref().clone();
+        dataset
+            .replace_field_metadata(vec![(
+                dataset.schema().field("emb").unwrap().id as u32,
+                HashMap::new(),
+            )])
+            .await
+            .unwrap();
+        let err = view.refresh().execute().await.unwrap_err().to_string();
+        assert!(err.contains("does not match binding 'fb_1'"), "{err}");
+    }
+
+    /// An input the view does not project is materialized on every refresh
+    /// path, before the provenance column, with the source's values.
+    #[tokio::test]
+    async fn test_internal_inputs_are_materialized_and_refreshed() {
+        use crate::materialized_view::tests::{computed_field, strict_people, test_binding};
+        let conn = connect("memory://").execute().await.unwrap();
+        let source = strict_people(&conn).await;
+        let mut prepared = crate::materialized_view::prepare_declaration(
+            &source,
+            Some(&[("id".to_string(), "id".to_string())]),
+            None,
+            None,
+        )
+        .await
+        .unwrap();
+        let input = prepared.input_column("name").unwrap();
+        let view = prepared
+            .with_computed_columns(
+                vec![(1, computed_field("emb", "fb_1", &input))],
+                &[test_binding("fb_1", &input, "emb")],
+            )
+            .unwrap()
+            .create("v")
+            .await
+            .unwrap();
+        let names: Vec<String> = view
+            .table()
+            .schema()
+            .await
+            .unwrap()
+            .fields()
+            .iter()
+            .map(|f| f.name().clone())
+            .collect();
+        assert_eq!(names, ["id", "emb", "__input_name", SOURCE_ROW_ID_COLUMN]);
+
+        let unfilled_inputs = || async {
+            view.table()
+                .count_rows(Some("__input_name IS NULL".to_string()))
+                .await
+                .unwrap()
+        };
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Rebuild
+        );
+        assert_eq!(view.table().count_rows(None).await.unwrap(), 3);
+        assert_eq!(unfilled_inputs().await, 0);
+
+        let more = arrow_array::RecordBatch::try_new(
+            source.schema().await.unwrap(),
+            vec![
+                Arc::new(Int32Array::from(vec![4])),
+                Arc::new(arrow_array::StringArray::from(vec!["d"])),
+            ],
+        )
+        .unwrap();
+        source.add(more).execute().await.unwrap();
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Incremental
+        );
+        assert_eq!(unfilled_inputs().await, 0);
+        assert_eq!(
+            view.table()
+                .count_rows(Some("__input_name = 'd'".to_string()))
+                .await
+                .unwrap(),
+            1
+        );
+
+        source
+            .update()
+            .column("name", "'z'")
+            .only_if("id = 1")
+            .execute()
+            .await
+            .unwrap();
+        view.refresh().execute().await.unwrap();
+        assert_eq!(
+            view.table()
+                .count_rows(Some("__input_name = 'z'".to_string()))
+                .await
+                .unwrap(),
+            1
+        );
+        assert_eq!(
+            unfilled(&view).await,
+            4,
+            "rewritten and new rows are unfilled"
+        );
+    }
+
+    /// A SQL declaration is filled by `refresh_column` on the view, which
+    /// commits a data replacement; the next refresh continues from its
+    /// watermark and keeps what the fill wrote, and only rows the view added
+    /// since come back unfilled.
+    #[tokio::test]
+    async fn test_a_sql_fill_is_not_drift() {
+        use crate::materialized_view::tests::{people, sql_field};
+        let conn = connect("memory://").execute().await.unwrap();
+        let source = people(&conn).await;
+        let view = crate::materialized_view::prepare_declaration(
+            &source,
+            Some(&[("id".to_string(), "id".to_string())]),
+            None,
+            None,
+        )
+        .await
+        .unwrap()
+        .with_computed_columns(
+            vec![(
+                1,
+                sql_field("next", arrow_schema::DataType::Int32, "id + 1", r#"["id"]"#),
+            )],
+            &[],
+        )
+        .unwrap()
+        .create("v")
+        .await
+        .unwrap();
+        let filled = || async {
+            view.table()
+                .count_rows(Some("next = id + 1".to_string()))
+                .await
+                .unwrap()
+        };
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Rebuild
+        );
+        assert_eq!(
+            view.table()
+                .refresh_column("next")
+                .await
+                .unwrap()
+                .rows_filled,
+            3
+        );
+        assert_eq!(filled().await, 3);
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+        assert_eq!(filled().await, 3);
+
+        append_people(&conn, vec![4], vec!["d"]).await;
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Incremental
+        );
+        assert_eq!(filled().await, 3);
+        assert_eq!(
+            view.table()
+                .refresh_column("next")
+                .await
+                .unwrap()
+                .rows_filled,
+            1
+        );
+        assert_eq!(filled().await, 4);
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+    }
+
+    /// A fill of a nested computed column writes its child fields; that is
+    /// still a fill, not drift.
+    #[tokio::test]
+    async fn test_a_nested_sql_fill_is_not_drift() {
+        use crate::materialized_view::tests::{people, sql_field};
+        let conn = connect("memory://").execute().await.unwrap();
+        let source = people(&conn).await;
+        let payload = sql_field(
+            "payload",
+            arrow_schema::DataType::Struct(
+                vec![arrow_schema::Field::new(
+                    "value",
+                    arrow_schema::DataType::Utf8,
+                    true,
+                )]
+                .into(),
+            ),
+            "named_struct('value', name)",
+            r#"["name"]"#,
+        );
+        let view = crate::materialized_view::prepare_declaration(
+            &source,
+            Some(&[("name".to_string(), "name".to_string())]),
+            None,
+            None,
+        )
+        .await
+        .unwrap()
+        .with_computed_columns(vec![(1, payload)], &[])
+        .unwrap()
+        .create("v")
+        .await
+        .unwrap();
+        view.refresh().execute().await.unwrap();
+        assert_eq!(
+            view.table()
+                .refresh_column("payload")
+                .await
+                .unwrap()
+                .rows_filled,
+            3
+        );
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+        assert_eq!(
+            view.table()
+                .count_rows(Some("payload.value = name".to_string()))
+                .await
+                .unwrap(),
+            3
+        );
+    }
 }

rust/lancedb/src/query.rs

@@ -1299,7 +1299,7 @@ impl VectorQuery {
     /// This can be useful when there is a narrow filter to allow these queries to
     /// spend more time searching and avoid potential false negatives.
     ///
-    /// Set to None to search all partitions, if needed, to satsify the limit
+    /// Set to None to search all partitions, if needed, to satisfy the limit
     pub fn maximum_nprobes(mut self, maximum_nprobes: Option<usize>) -> Result<Self> {
         if let Some(maximum_nprobes) = maximum_nprobes {
             if maximum_nprobes == 0 {

rust/lancedb/src/remote/table.rs

@@ -8021,7 +8021,7 @@ mod tests {
             match request.url().path() {
                 "/v1/table/my_table/backfill_column" => http::Response::builder()
                     .status(202)
-                    .body(r#"{"job_id": "j-42"}"#.as_bytes().to_vec())
+                    .body(br#"{"job_id": "j-42"}"#.to_vec())
                     .unwrap(),
                 "/v1/jobs/describe" => http::Response::builder()
                     .status(200)

rust/lancedb/src/table.rs

@@ -240,7 +240,7 @@ enum BadVectorHandling {
     /// An error is returned
     #[default]
     Error,
-    /// The offending row is droppped
+    /// The offending row is dropped
     Drop,
     /// The invalid/missing items are replaced by fill_value
     Fill(f32),
@@ -1326,7 +1326,7 @@ impl Table {
     /// Note: if your condition is something like "some_id_column == 7" and
     /// you are updating many rows (with different ids) then you will get
     /// better performance with a single [`merge_insert`] call instead of
-    /// repeatedly calilng this method.
+    /// repeatedly calling this method.
     pub fn update(&self) -> UpdateBuilder {
         UpdateBuilder::new(self.inner.clone())
     }
@@ -2804,7 +2804,7 @@ impl NativeTable {
         namespace_client: Option<Arc<dyn LanceNamespace>>,
         pushdown_operations: HashSet<NamespaceClientPushdownOperation>,
     ) -> Result<Self> {
-        computed_columns::ensure_no_foreign_declarations(batches.arrow_schema().fields())?;
+        let batches = computed_columns::admit_create_source(batches)?;
         // Default params uses format v1.
         let params = params.unwrap_or(WriteParams {
             ..Default::default()
@@ -2904,6 +2904,7 @@ impl NativeTable {
         pushdown_operations: HashSet<NamespaceClientPushdownOperation>,
         session: Option<Arc<lance::session::Session>>,
     ) -> Result<Self> {
+        let batches = computed_columns::admit_create_source(batches)?;
         // Build table_id from namespace + name for the storage options provider
         let mut table_id = namespace.clone();
         table_id.push(name.to_string());
@@ -5677,7 +5678,7 @@ mod tests {
             TableStatistics {
                 num_rows: 250,
                 num_indices: 0,
-                total_bytes: 8925,
+                total_bytes: 8969,
                 fragment_stats: FragmentStatistics {
                     num_fragments: 11,
                     num_small_fragments: 11,

rust/lancedb/src/table/computed_columns.rs

@@ -21,6 +21,7 @@
 //! [`computed_columns`] and [`computed_column_from_field`] read declarations
 //! back off a schema.
 
+use futures::StreamExt;
 use std::collections::{BTreeSet, HashMap, HashSet};
 use std::sync::Arc;
 
@@ -1338,6 +1339,106 @@ pub(crate) fn ensure_batch_writes_no_computed_values(
     Ok(())
 }
 
+/// Validate every computed-column declaration `schema` carries against the
+/// schema itself: every field with declaration metadata is a complete
+/// declaration, a SQL declaration re-plans to the field it declares, a
+/// Function declaration satisfies the binding contract, and no declaration
+/// reads another computed column. What passes here is what `refresh_column`
+/// can execute.
+pub(crate) fn ensure_declarations_are_planned(schema: &ArrowSchema) -> Result<()> {
+    let invalid = |message: String| Error::InvalidInput { message };
+    // A field with any declaration key is a declaration; a partial one is
+    // not "no declaration", it is a broken one.
+    for field in schema.fields() {
+        if field.metadata().keys().any(|k| is_declaration_key(k))
+            && computed_column_from_field(field).is_none()
+        {
+            return Err(invalid(format!(
+                "field '{}' carries an incomplete computed-column declaration",
+                field.name()
+            )));
+        }
+    }
+    let declared: HashSet<String> = computed_columns(schema)
+        .into_iter()
+        .map(|c| c.name)
+        .collect();
+    for column in computed_columns(schema) {
+        let field = schema.field_with_name(&column.name)?;
+        if !field.is_nullable() {
+            return Err(invalid(format!(
+                "computed column '{}' must be nullable until a refresh fills it",
+                column.name
+            )));
+        }
+        match &column.kind {
+            ComputedColumnKind::Sql { expression } => {
+                let others: Vec<ArrowField> = schema
+                    .fields()
+                    .iter()
+                    .filter(|f| f.name() != &column.name)
+                    .map(|f| f.as_ref().clone())
+                    .collect();
+                let bound = bind(Arc::new(ArrowSchema::new(others)), &column.name, expression)?;
+                if let Some(input) = bound.roots.iter().find(|r| declared.contains(*r)) {
+                    return Err(invalid(format!(
+                        "computed column '{}' reads computed column '{input}'",
+                        column.name
+                    )));
+                }
+                if &bound.data_type != field.data_type() {
+                    return Err(invalid(format!(
+                        "computed column '{}' is declared as {} but its expression yields {}",
+                        column.name,
+                        field.data_type(),
+                        bound.data_type
+                    )));
+                }
+                let mut declared_inputs = column.inputs.clone();
+                declared_inputs.sort();
+                if declared_inputs != bound.inputs {
+                    return Err(invalid(format!(
+                        "computed column '{}' declares inputs {:?} but its expression reads {:?}",
+                        column.name, declared_inputs, bound.inputs
+                    )));
+                }
+            }
+            ComputedColumnKind::Function { binding_id, .. } => {
+                // The binding validator resolves each input's leaf; the
+                // no-computed-input rule is about the root it hangs from.
+                let bindings = function_bindings(schema)?;
+                let Some(binding) = bindings.iter().find(|b| b.binding_id() == binding_id) else {
+                    continue; // reported by the binding validator below
+                };
+                // Roots come from the canonical path parser: a quoted
+                // top-level name may itself contain a dot.
+                if let Some(input) = binding
+                    .inputs()
+                    .iter()
+                    .filter_map(|input| resolve_field_path(schema, &input.field_path).ok())
+                    .map(|resolved| resolved.root.name().as_str())
+                    .find(|r| declared.contains(*r))
+                {
+                    return Err(invalid(format!(
+                        "computed column '{}' reads computed column '{input}'",
+                        column.name
+                    )));
+                }
+            }
+            ComputedColumnKind::Unrecognized { kind } => {
+                return Err(Error::NotSupported {
+                    message: format!(
+                        "computed column '{}' is defined by '{kind}', which this version \
+                         of lancedb cannot fill",
+                        column.name
+                    ),
+                });
+            }
+        }
+    }
+    ensure_supported_function_metadata(schema)
+}
+
 /// Reject fields carrying declaration metadata that did not come through
 /// [`plan`]. One authority for creation, overwrite and raw transforms.
 pub(crate) fn ensure_no_foreign_declarations<'a>(
@@ -1796,6 +1897,54 @@ pub(super) async fn add_foreign_kind(table: &crate::Table, name: &str, kind: &st
     .unwrap();
 }
 
+/// Admit a table's initial data: every declaration it carries is validated,
+/// and the stream refuses any batch with values in a computed column, whose
+/// values come from refresh alone. One boundary for every way a table is
+/// created.
+pub(crate) fn admit_create_source<S: lance_datafusion::utils::StreamingWriteSource>(
+    batches: S,
+) -> Result<UnfilledDeclarations<S>> {
+    let schema = batches.arrow_schema();
+    ensure_declarations_are_planned(&schema)?;
+    let declared = computed_columns(&schema)
+        .into_iter()
+        .map(|c| c.name)
+        .collect();
+    Ok(UnfilledDeclarations {
+        inner: batches,
+        declared,
+    })
+}
+
+/// A write source whose computed columns must arrive unfilled.
+pub(crate) struct UnfilledDeclarations<S> {
+    inner: S,
+    declared: Vec<String>,
+}
+
+impl<S: lance_datafusion::utils::StreamingWriteSource> lance_datafusion::utils::StreamingWriteSource
+    for UnfilledDeclarations<S>
+{
+    fn arrow_schema(&self) -> SchemaRef {
+        self.inner.arrow_schema()
+    }
+
+    fn into_stream(self) -> datafusion_physical_plan::SendableRecordBatchStream {
+        if self.declared.is_empty() {
+            return self.inner.into_stream();
+        }
+        let schema = self.inner.arrow_schema();
+        let declared = self.declared;
+        let stream = self.inner.into_stream().map(move |batch| {
+            let batch = batch?;
+            ensure_batch_writes_no_computed_values(&declared, &batch)
+                .map_err(|e| datafusion_common::DataFusionError::External(Box::new(e)))?;
+            Ok(batch)
+        });
+        Box::pin(datafusion_physical_plan::stream::RecordBatchStreamAdapter::new(schema, stream))
+    }
+}
+
 #[cfg(test)]
 mod tests {
     /// The gate's reproducer: the validator applies the same schema-level
@@ -2646,6 +2795,8 @@ mod tests {
         );
     }
 
+    /// A create carries a declaration only if it re-plans completely; this
+    /// one lacks its inputs and is refused before its forged value matters.
     #[tokio::test]
     async fn test_create_table_cannot_inject_a_declaration() {
         let conn = connect("memory://").execute().await.unwrap();
@@ -2673,7 +2824,7 @@ mod tests {
             .await
             .unwrap_err();
         assert!(
-            matches!(&err, Error::InvalidInput { message } if message.contains("computed()")),
+            matches!(&err, Error::InvalidInput { message } if message.contains("computed column 'doubled'")),
             "{err:?}"
         );
     }

rust/lancedb/src/table/datafusion/blob_coerce.rs

@@ -5,12 +5,17 @@
 //!
 //! [`super::cast::cast_to_table_schema`] calls [`coerce_blob_expr`].
 
+use std::fmt;
+use std::hash::{Hash, Hasher};
 use std::sync::Arc;
 
-use arrow_schema::{DataType, Field, FieldRef, Fields};
+use arrow_array::{Array, BooleanArray, RecordBatch};
+use arrow_schema::{DataType, Field, FieldRef, Fields, Schema};
+use arrow_select::nullif::nullif;
 use datafusion::functions::core::{get_field, named_struct};
 use datafusion_common::ScalarValue;
 use datafusion_common::config::ConfigOptions;
+use datafusion_expr::ColumnarValue;
 use datafusion_physical_expr::ScalarFunctionExpr;
 use datafusion_physical_expr::expressions::{CastExpr, Literal};
 use datafusion_physical_plan::PhysicalExpr;
@@ -133,16 +138,102 @@ pub(super) fn coerce_blob_expr(
         ns_args.push(value);
     }
 
-    let expr: Arc<dyn PhysicalExpr> = Arc::new(ScalarFunctionExpr::new(
+    let built: Arc<dyn PhysicalExpr> = Arc::new(ScalarFunctionExpr::new(
         &format!("named_struct({})", table_field.name()),
         named_struct(),
         ns_args,
         table_field.clone(),
         config.clone(),
     ));
+
+    // `named_struct` always yields a valid struct, so a null input would land
+    // as a row that set neither `data` nor `uri` -- not an absent blob but a
+    // malformed one, which Lance rejects on write.
+    let expr: Arc<dyn PhysicalExpr> = Arc::new(AbsentBlobIsNull {
+        source: input_expr,
+        built,
+        field: table_field.clone(),
+    });
     Ok((expr, table_field.clone()))
 }
 
+/// Carries the source column's nullity onto the struct built for it.
+///
+/// This is its own expression rather than a `CASE` because the projection
+/// takes its output field from `return_field`, and the generic implementation
+/// rebuilds a bare field -- which would drop the `lance.blob.v2` extension
+/// metadata and stop the column being recognised as a blob at all.
+#[derive(Debug, Clone)]
+struct AbsentBlobIsNull {
+    source: Arc<dyn PhysicalExpr>,
+    built: Arc<dyn PhysicalExpr>,
+    field: FieldRef,
+}
+
+impl fmt::Display for AbsentBlobIsNull {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "absent_blob_is_null({}, {})", self.source, self.built)
+    }
+}
+
+impl PartialEq for AbsentBlobIsNull {
+    fn eq(&self, other: &Self) -> bool {
+        self.source.eq(&other.source) && self.built.eq(&other.built) && self.field == other.field
+    }
+}
+
+impl Eq for AbsentBlobIsNull {}
+
+impl Hash for AbsentBlobIsNull {
+    fn hash<H: Hasher>(&self, state: &mut H) {
+        self.source.hash(state);
+        self.built.hash(state);
+        self.field.hash(state);
+    }
+}
+
+impl PhysicalExpr for AbsentBlobIsNull {
+    fn return_field(&self, _input_schema: &Schema) -> datafusion_common::Result<FieldRef> {
+        Ok(self.field.clone())
+    }
+
+    fn nullable(&self, _input_schema: &Schema) -> datafusion_common::Result<bool> {
+        Ok(true)
+    }
+
+    fn evaluate(&self, batch: &RecordBatch) -> datafusion_common::Result<ColumnarValue> {
+        let rows = batch.num_rows();
+        let built = self.built.evaluate(batch)?.into_array(rows)?;
+        let source = self.source.evaluate(batch)?.into_array(rows)?;
+        let Some(nulls) = source.logical_nulls() else {
+            return Ok(ColumnarValue::Array(built));
+        };
+        // `nullif` nulls the rows the mask marks true, which is where the
+        // source had no value.
+        let absent = BooleanArray::new(!nulls.inner(), None);
+        Ok(ColumnarValue::Array(nullif(built.as_ref(), &absent)?))
+    }
+
+    fn children(&self) -> Vec<&Arc<dyn PhysicalExpr>> {
+        vec![&self.source, &self.built]
+    }
+
+    fn with_new_children(
+        self: Arc<Self>,
+        children: Vec<Arc<dyn PhysicalExpr>>,
+    ) -> datafusion_common::Result<Arc<dyn PhysicalExpr>> {
+        Ok(Arc::new(Self {
+            source: children[0].clone(),
+            built: children[1].clone(),
+            field: self.field.clone(),
+        }))
+    }
+
+    fn fmt_sql(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "{self}")
+    }
+}
+
 enum BlobInputShape<'a> {
     Bytes,
     String,
@@ -313,6 +404,11 @@ mod tests {
         let data = image.column_by_name("data").unwrap();
         assert!(!data.is_null(0));
         assert!(data.is_null(1));
+        // The row itself has to be null, not merely a struct whose children
+        // are. A present-but-empty struct set neither `data` nor `uri`, which
+        // Lance rejects as malformed rather than reading as an absent blob.
+        assert!(!image.is_null(0));
+        assert!(image.is_null(1));
     }
 
     #[tokio::test]

rust/lancedb/src/table/dataset.rs

@@ -52,7 +52,7 @@ enum ConsistencyMode {
     /// refresh_window = min(3s, TTL/4)
     ///
     /// | t < TTL - refresh_window | t < TTL                           | t >= TTL            |
-    /// |  Return value            | Background refresh & return value |  syncronous refresh |
+    /// |  Return value            | Background refresh & return value |  synchronous refresh |
     Eventual(BackgroundCache<Arc<Dataset>, Error>),
 }
 

rust/lancedb/src/table/merge.rs

@@ -103,7 +103,7 @@ impl MergeInsertBuilder {
     /// but that behavior is subject to change.
     ///
     /// An optional condition may be specified.  If it is, then only
-    /// matched rows that satisfy the condtion will be updated.  Any
+    /// matched rows that satisfy the condition will be updated.  Any
     /// rows that do not satisfy the condition will be left as they
     /// are.  Failing to satisfy the condition does not cause a
     /// "matched row" to become a "not matched" row.

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

@@ -904,7 +904,7 @@ fn unsharded_shard_id() -> Uuid {
 
 /// Build a [`ShardWriterConfig`] from the persisted `writer_config_defaults`.
 ///
-/// Unknown or unparseable keys are ignored; absent keys keep the
+/// Unknown or unparsable keys are ignored; absent keys keep the
 /// [`ShardWriterConfig`] default. The shard id is set by `mem_wal_writer`.
 fn shard_writer_config_from_defaults(defaults: &HashMap<String, String>) -> ShardWriterConfig {
     let mut config = ShardWriterConfig::default().with_shard_spec_id(SHARDING_SPEC_ID);

rust/lancedb/tests/blob_integration.rs

@@ -19,6 +19,7 @@ use lancedb::{
     connect, connect_namespace,
     database::listing::{
         ListingDatabaseOptions, NewTableConfig, OPT_NEW_TABLE_ENABLE_STABLE_ROW_IDS,
+        OPT_NEW_TABLE_STORAGE_VERSION,
     },
     query::{ExecutableQuery, QueryBase},
     table::{AddDataMode, CompactionOptions, OptimizeAction, OptimizeStats, WriteOptions},
@@ -146,7 +147,10 @@ async fn non_blob_table_keeps_default_format_and_row_id_setting() -> Result<()>
     let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int64, false)]));
     let table = db.create_empty_table("t", schema).execute().await?;
 
-    assert!(!supports_blob_v2(storage_format_version(&table).await));
+    assert_eq!(
+        storage_format_version(&table).await,
+        LanceFileVersion::Stable.resolve()
+    );
     assert!(!uses_stable_row_ids(&table).await);
     Ok(())
 }
@@ -809,7 +813,11 @@ async fn fetch_blobs_rejects_unknown_column() -> Result<()> {
 #[tokio::test]
 async fn fetch_blobs_rejects_legacy_v1_blob_column() -> Result<()> {
     let tmp = tempdir().unwrap();
-    let db = connect(tmp.path().to_str().unwrap()).execute().await?;
+    // Legacy v1 blob columns are only writable at file version <= 2.1.
+    let db = connect(tmp.path().to_str().unwrap())
+        .storage_options([(OPT_NEW_TABLE_STORAGE_VERSION, "2.1")])
+        .execute()
+        .await?;
     let legacy = Field::new("image", DataType::LargeBinary, true).with_metadata(
         std::collections::HashMap::from([("lance-encoding:blob".to_string(), "true".to_string())]),
     );