chore: upgrade lance to 0.9.16 (#975)

When we turned on fat LTO builds, we made the release build job **much**
more compute and memory intensive. The ARM runners have particularly low
memory per core, which makes them susceptible to OOM errors. To avoid
issues, I have enabled memory swap on ARM and bumped the side of the
runner.

.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.4.4
+current_version = 0.4.10
 commit = True
 message = Bump version: {current_version} → {new_version}
 tag = True

.cargo/config.toml

@@ -0,0 +1,35 @@
+[profile.release]
+lto = "fat"
+codegen-units = 1
+
+[profile.release-with-debug]
+inherits = "release"
+debug = true
+# Prioritize compile time over runtime performance
+codegen-units = 16
+lto = "thin"
+
+[target.'cfg(all())']
+rustflags = [
+    "-Wclippy::all",
+    "-Wclippy::style",
+    "-Wclippy::fallible_impl_from",
+    "-Wclippy::manual_let_else",
+    "-Wclippy::redundant_pub_crate",
+    "-Wclippy::string_add_assign",
+    "-Wclippy::string_add",
+    "-Wclippy::string_lit_as_bytes",
+    "-Wclippy::string_to_string",
+    "-Wclippy::use_self",
+    "-Dclippy::cargo",
+    "-Dclippy::dbg_macro",
+    # not too much we can do to avoid multiple crate versions
+    "-Aclippy::multiple-crate-versions",
+    "-Aclippy::wildcard_dependencies",
+]
+
+[target.x86_64-unknown-linux-gnu]
+rustflags = ["-C", "target-cpu=haswell", "-C", "target-feature=+avx2,+fma,+f16c"]
+
+[target.aarch64-apple-darwin]
+rustflags = ["-C", "target-cpu=apple-m1", "-C", "target-feature=+neon,+fp16,+fhm,+dotprod"]

.github/workflows/cargo-publish.yml

@@ -16,7 +16,7 @@ jobs:
     # Only runs on tags that matches the make-release action
     if: startsWith(github.ref, 'refs/tags/v')
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - uses: Swatinem/rust-cache@v2
         with:
           workspaces: rust

.github/workflows/docs.yml

@@ -27,9 +27,9 @@ jobs:
     runs-on: ubuntu-22.04
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Set up Python
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: "3.10"
           cache: "pip"
@@ -42,7 +42,7 @@ jobs:
       - name: Set up node
         uses: actions/setup-node@v3
         with:
-          node-version: ${{ matrix.node-version }}
+          node-version: 20
           cache: 'npm'
           cache-dependency-path: node/package-lock.json
       - uses: Swatinem/rust-cache@v2
@@ -62,8 +62,9 @@ jobs:
         run: |
           npx typedoc --plugin typedoc-plugin-markdown --out ../docs/src/javascript src/index.ts
       - name: Build docs
+        working-directory: docs
         run: |
-          PYTHONPATH=. mkdocs build -f docs/mkdocs.yml
+          PYTHONPATH=. mkdocs build
       - name: Setup Pages
         uses: actions/configure-pages@v2
       - name: Upload artifact
@@ -72,4 +73,4 @@ jobs:
           path: "docs/site"
       - name: Deploy to GitHub Pages
         id: deployment
-        uses: actions/deploy-pages@v1
+        uses: actions/deploy-pages@v1

.github/workflows/docs_test.yml

@@ -18,24 +18,20 @@ on:
 env:
   # Disable full debug symbol generation to speed up CI build and keep memory down
   # "1" means line tables only, which is useful for panic tracebacks.
-  RUSTFLAGS: "-C debuginfo=1"
+  RUSTFLAGS: "-C debuginfo=1 -C target-cpu=native -C target-feature=+f16c,+avx2,+fma"
   RUST_BACKTRACE: "1"
 
 jobs:
   test-python:
     name: Test doc python code
-    runs-on: ${{ matrix.os }}
-    strategy:
-      matrix:
-        python-minor-version: [ "11" ]
-        os: ["ubuntu-22.04"]
+    runs-on: "ubuntu-latest"
     steps:
     - name: Checkout
-      uses: actions/checkout@v3
+      uses: actions/checkout@v4
     - name: Set up Python
-      uses: actions/setup-python@v4
+      uses: actions/setup-python@v5
       with:
-        python-version: 3.${{ matrix.python-minor-version }}
+        python-version: 3.11
         cache: "pip"
         cache-dependency-path: "docs/test/requirements.txt"
     - name: Build Python
@@ -52,45 +48,42 @@ jobs:
         for d in *; do cd "$d"; echo "$d".py; python "$d".py; cd ..; done
   test-node:
     name: Test doc nodejs code
-    runs-on: ${{ matrix.os }}
+    runs-on: "ubuntu-latest"
+    timeout-minutes: 45
     strategy:
-      matrix:
-        node-version: [ "18" ]
-        os: ["ubuntu-22.04"]
+      fail-fast: false
     steps:
     - name: Checkout
-      uses: actions/checkout@v3
+      uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true
     - name: Set up Node
-      uses: actions/setup-node@v3
+      uses: actions/setup-node@v4
       with:
-        node-version: ${{ matrix.node-version }}
+        node-version: 20
     - name: Install dependecies needed for ubuntu
-      if: ${{ matrix.os == 'ubuntu-22.04' }}
       run: |
         sudo apt install -y protobuf-compiler libssl-dev
-    - name: Install node dependencies
-      run: |
-        cd docs/test
-        npm install 
     - name: Rust cache
       uses: swatinem/rust-cache@v2
-    - name: Install LanceDB
+    - name: Install node dependencies
       run: |
-        cd docs/test/node_modules/vectordb 
+        sudo swapoff -a
+        sudo fallocate -l 8G /swapfile
+        sudo chmod 600 /swapfile
+        sudo mkswap /swapfile
+        sudo swapon /swapfile
+        sudo swapon --show
+        cd node
         npm ci
         npm run build-release
-        npm run tsc
-    - name: Create test files
-      run: |
-        cd docs/test
-        node md_testing.js
+        cd ../docs
+        npm install
     - name: Test
       env:
         LANCEDB_URI: ${{ secrets.LANCEDB_URI }}
         LANCEDB_DEV_API_KEY: ${{ secrets.LANCEDB_DEV_API_KEY }}
       run: |
-        cd docs/test/node
-        for d in *; do cd "$d"; echo "$d".js; node "$d".js; cd ..; done
+        cd docs
+        npm t

.github/workflows/make-release-commit.yml

@@ -26,7 +26,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Check out main
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           ref: main
           persist-credentials: false
@@ -37,10 +37,10 @@ jobs:
         run: |
           git config user.name 'Lance Release'
           git config user.email 'lance-dev@lancedb.com'
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v4
+      - name: Set up Python 3.11
+        uses: actions/setup-python@v5
         with:
-          python-version: "3.10"
+          python-version: "3.11"
       - name: Bump version, create tag and commit
         run: |
           pip install bump2version

.github/workflows/node.yml

@@ -32,7 +32,7 @@ jobs:
         shell: bash
         working-directory: node
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true
@@ -57,7 +57,7 @@ jobs:
         shell: bash
         working-directory: node
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true
@@ -89,7 +89,7 @@ jobs:
         shell: bash
         working-directory: node
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true
@@ -128,7 +128,7 @@ jobs:
       # this one is for dynamodb
       DYNAMODB_ENDPOINT: http://localhost:4566
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true

.github/workflows/nodejs.yml

@@ -29,7 +29,7 @@ jobs:
         shell: bash
         working-directory: nodejs
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true
@@ -61,7 +61,7 @@ jobs:
         shell: bash
         working-directory: nodejs
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true
@@ -84,13 +84,13 @@ jobs:
       run: npm run test
   macos:
     timeout-minutes: 30
-    runs-on: "macos-13"
+    runs-on: "macos-14"
     defaults:
       run:
         shell: bash
         working-directory: nodejs
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true

.github/workflows/npm-publish.yml

@@ -15,7 +15,7 @@ jobs:
         working-directory: node
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - uses: actions/setup-node@v3
         with:
           node-version: 20
@@ -45,13 +45,13 @@ jobs:
             runner: macos-13
           - arch: aarch64-apple-darwin
             # xlarge is implicitly arm64.
-            runner: macos-13-xlarge
+            runner: macos-14
     runs-on: ${{ matrix.config.runner }}
     # Only runs on tags that matches the make-release action
     if: startsWith(github.ref, 'refs/tags/v')
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Install system dependencies
         run: brew install protobuf
       - name: Install npm dependencies
@@ -66,7 +66,7 @@ jobs:
           name: native-darwin
           path: |
             node/dist/lancedb-vectordb-darwin*.tgz
-         
+
 
   node-linux:
     name: node-linux (${{ matrix.config.arch}}-unknown-linux-gnu
@@ -80,10 +80,25 @@ jobs:
           - arch: x86_64
             runner: ubuntu-latest
           - arch: aarch64
-            runner: buildjet-4vcpu-ubuntu-2204-arm
+            # For successful fat LTO builds, we need a large runner to avoid OOM errors.
+            runner: buildjet-16vcpu-ubuntu-2204-arm
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
+      # Buildjet aarch64 runners have only 1.5 GB RAM per core, vs 3.5 GB per core for
+      # x86_64 runners. To avoid OOM errors on ARM, we create a swap file.
+      - name: Configure aarch64 build
+        if: ${{ matrix.config.arch == 'aarch64' }}
+        run: |
+          free -h
+          sudo fallocate -l 16G /swapfile
+          sudo chmod 600 /swapfile
+          sudo mkswap /swapfile
+          sudo swapon /swapfile
+          echo "/swapfile swap swap defaults 0 0" >> sudo /etc/fstab
+          # print info
+          swapon --show
+          free -h
       - name: Build Linux Artifacts
         run: |
           bash ci/build_linux_artifacts.sh ${{ matrix.config.arch }}
@@ -104,7 +119,7 @@ jobs:
         target: [x86_64-pc-windows-msvc]
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Install Protoc v21.12
         working-directory: C:\
         run: |
@@ -154,7 +169,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           ref: main
           persist-credentials: false

.github/workflows/pypi-publish.yml

@@ -14,9 +14,9 @@ jobs:
         shell: bash
         working-directory: python
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - name: Set up Python
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: "3.8"
       - name: Build distribution

.github/workflows/python-make-release-commit.yml

@@ -26,7 +26,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - name: Check out main
-      uses: actions/checkout@v3
+      uses: actions/checkout@v4
       with:
         ref: main
         persist-credentials: false
@@ -37,10 +37,10 @@ jobs:
       run: |
         git config user.name 'Lance Release'
         git config user.email 'lance-dev@lancedb.com'
-    - name: Set up Python 3.10
-      uses: actions/setup-python@v4
+    - name: Set up Python
+      uses: actions/setup-python@v5
       with:
-        python-version: "3.10"
+        python-version: "3.11"
     - name: Bump version, create tag and commit
       working-directory: python
       run: |

.github/workflows/python.yml

@@ -18,19 +18,19 @@ jobs:
     timeout-minutes: 30
     strategy:
       matrix:
-        python-minor-version: [ "8", "9", "10", "11" ]
+        python-minor-version: [ "8", "11" ]
     runs-on: "ubuntu-22.04"
     defaults:
       run:
         shell: bash
         working-directory: python
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true
     - name: Set up Python
-      uses: actions/setup-python@v4
+      uses: actions/setup-python@v5
       with:
         python-version: 3.${{ matrix.python-minor-version }}
     - name: Install lancedb
@@ -55,7 +55,7 @@ jobs:
           - name: x86 Mac
             runner: macos-13
           - name: Arm Mac
-            runner: macos-13-xlarge
+            runner: macos-14
           - name: x86 Windows
             runner: windows-latest
     runs-on: "${{ matrix.config.runner }}"
@@ -64,12 +64,12 @@ jobs:
         shell: bash
         working-directory: python
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true
     - name: Set up Python
-      uses: actions/setup-python@v4
+      uses: actions/setup-python@v5
       with:
         python-version: "3.11"
     - name: Install lancedb
@@ -87,12 +87,12 @@ jobs:
         shell: bash
         working-directory: python
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         fetch-depth: 0
         lfs: true
     - name: Set up Python
-      uses: actions/setup-python@v4
+      uses: actions/setup-python@v5
       with:
         python-version: 3.9
     - name: Install lancedb

.github/workflows/rust.yml

@@ -32,7 +32,7 @@ jobs:
         shell: bash
         working-directory: rust
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
           fetch-depth: 0
           lfs: true
@@ -55,7 +55,7 @@ jobs:
         shell: bash
         working-directory: rust
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
           fetch-depth: 0
           lfs: true
@@ -70,18 +70,20 @@ jobs:
       run: cargo build --all-features
     - name: Run tests
       run: cargo test --all-features
+    - name: Run examples
+      run: cargo run --example simple
   macos:
     timeout-minutes: 30
     strategy:
       matrix:
-        mac-runner: [ "macos-13", "macos-13-xlarge" ]
+        mac-runner: [ "macos-13", "macos-14" ]
     runs-on: "${{ matrix.mac-runner }}"
     defaults:
       run:
         shell: bash
         working-directory: rust
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
         with:
             fetch-depth: 0
             lfs: true
@@ -99,7 +101,7 @@ jobs:
   windows:
     runs-on: windows-2022
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - uses: Swatinem/rust-cache@v2
         with:
           workspaces: rust

.github/workflows/update_package_lock_run.yml

@@ -8,7 +8,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           ref: main
           persist-credentials: false

Cargo.toml

@@ -6,24 +6,27 @@ resolver = "2"
 
 [workspace.package]
 edition = "2021"
-authors = ["Lance Devs <dev@lancedb.com>"]
+authors = ["LanceDB Devs <dev@lancedb.com>"]
 license = "Apache-2.0"
 repository = "https://github.com/lancedb/lancedb"
+description = "Serverless, low-latency vector database for AI applications"
+keywords = ["lancedb", "lance", "database", "vector", "search"]
+categories = ["database-implementations"]
 
 [workspace.dependencies]
-lance = { "version" = "=0.9.9", "features" = ["dynamodb"] }
-lance-index = { "version" = "=0.9.9" }
-lance-linalg = { "version" = "=0.9.9" }
-lance-testing = { "version" = "=0.9.9" }
+lance = { "version" = "=0.9.16", "features" = ["dynamodb"] }
+lance-index = { "version" = "=0.9.16" }
+lance-linalg = { "version" = "=0.9.16" }
+lance-testing = { "version" = "=0.9.16" }
 # Note that this one does not include pyarrow
-arrow = { version = "49.0.0", optional = false }
-arrow-array = "49.0"
-arrow-data = "49.0"
-arrow-ipc = "49.0"
-arrow-ord = "49.0"
-arrow-schema = "49.0"
-arrow-arith = "49.0"
-arrow-cast = "49.0"
+arrow = { version = "50.0", optional = false }
+arrow-array = "50.0"
+arrow-data = "50.0"
+arrow-ipc = "50.0"
+arrow-ord = "50.0"
+arrow-schema = "50.0"
+arrow-arith = "50.0"
+arrow-cast = "50.0"
 async-trait = "0"
 chrono = "0.4.23"
 half = { "version" = "=2.3.1", default-features = false, features = [

README.md

@@ -51,12 +51,19 @@ npm install vectordb
 const lancedb = require('vectordb');
 const db = await lancedb.connect('data/sample-lancedb');
 
-const table = await db.createTable('vectors',
-      [{ id: 1, vector: [0.1, 0.2], item: "foo", price: 10 },
-       { id: 2, vector: [1.1, 1.2], item: "bar", price: 50 }])
+const table = await db.createTable({
+  name: 'vectors',
+  data:  [
+    { id: 1, vector: [0.1, 0.2], item: "foo", price: 10 },
+    { id: 2, vector: [1.1, 1.2], item: "bar", price: 50 }
+  ]
+})
 
 const query = table.search([0.1, 0.3]).limit(2);
 const results = await query.execute();
+
+// You can also search for rows by specific criteria without involving a vector search.
+const rowsByCriteria = await table.search(undefined).where("price >= 10").execute();
 ```
 
 **Python**

ci/build_linux_artifacts.sh

@@ -13,7 +13,9 @@ docker build \
     .
 popd
 
+# We turn on memory swap to avoid OOM killer
 docker run \
     -v $(pwd):/io -w /io \
+    --memory-swap=-1 \
     lancedb-node-manylinux \
     bash ci/manylinux_node/build.sh $ARCH

docs/README.md

@@ -33,3 +33,12 @@ You can run a local server to test the docs prior to deployment by navigating to
 cd docs
 mkdocs serve
 ```
+
+### Run doctest for typescript example
+
+```bash
+cd lancedb/docs
+npm i
+npm run build
+npm run all
+```

docs/mkdocs.yml

@@ -67,7 +67,9 @@ markdown_extensions:
     line_spans: __span
     pygments_lang_class: true
 - pymdownx.inlinehilite
-- pymdownx.snippets
+- pymdownx.snippets:
+    base_path: ..
+    dedent_subsections: true
 - pymdownx.superfences
 - pymdownx.tabbed:
     alternate_style: true
@@ -88,6 +90,9 @@ nav:
     - Building an ANN index: ann_indexes.md
     - Vector Search: search.md
     - Full-text search: fts.md
+    - Hybrid search:
+      - Overview: hybrid_search/hybrid_search.md
+      - Airbnb financial data example: notebooks/hybrid_search.ipynb
     - Filtering: sql.md
     - Versioning & Reproducibility: notebooks/reproducibility.ipynb
     - Configuring Storage: guides/storage.md
@@ -130,6 +135,7 @@ nav:
   - ⚙️ API reference:
     - 🐍 Python: python/python.md
     - 👾 JavaScript: javascript/modules.md
+    - 🦀 Rust: https://docs.rs/vectordb/latest/vectordb/
   - ☁️ LanceDB Cloud:
     - Overview: cloud/index.md
     - API reference:
@@ -148,6 +154,9 @@ nav:
     - Building an ANN index: ann_indexes.md
     - Vector Search: search.md
     - Full-text search: fts.md
+    - Hybrid search:
+      - Overview: hybrid_search/hybrid_search.md
+      - Airbnb financial data example: notebooks/hybrid_search.ipynb
     - Filtering: sql.md
     - Versioning & Reproducibility: notebooks/reproducibility.ipynb
     - Configuring Storage: guides/storage.md
@@ -195,6 +204,9 @@ extra_css:
   - styles/global.css
   - styles/extra.css
 
+extra_javascript:
+  - "extra_js/init_ask_ai_widget.js"
+
 extra:
   analytics:
     provider: google

docs/package-lock.json

@@ -0,0 +1,132 @@
+{
+  "name": "lancedb-docs-test",
+  "version": "1.0.0",
+  "lockfileVersion": 3,
+  "requires": true,
+  "packages": {
+    "": {
+      "name": "lancedb-docs-test",
+      "version": "1.0.0",
+      "license": "Apache 2",
+      "dependencies": {
+        "apache-arrow": "file:../node/node_modules/apache-arrow",
+        "vectordb": "file:../node"
+      },
+      "devDependencies": {
+        "@types/node": "^20.11.8",
+        "typescript": "^5.3.3"
+      }
+    },
+    "../node": {
+      "name": "vectordb",
+      "version": "0.4.6",
+      "cpu": [
+        "x64",
+        "arm64"
+      ],
+      "license": "Apache-2.0",
+      "os": [
+        "darwin",
+        "linux",
+        "win32"
+      ],
+      "dependencies": {
+        "@apache-arrow/ts": "^14.0.2",
+        "@neon-rs/load": "^0.0.74",
+        "apache-arrow": "^14.0.2",
+        "axios": "^1.4.0"
+      },
+      "devDependencies": {
+        "@neon-rs/cli": "^0.0.160",
+        "@types/chai": "^4.3.4",
+        "@types/chai-as-promised": "^7.1.5",
+        "@types/mocha": "^10.0.1",
+        "@types/node": "^18.16.2",
+        "@types/sinon": "^10.0.15",
+        "@types/temp": "^0.9.1",
+        "@types/uuid": "^9.0.3",
+        "@typescript-eslint/eslint-plugin": "^5.59.1",
+        "cargo-cp-artifact": "^0.1",
+        "chai": "^4.3.7",
+        "chai-as-promised": "^7.1.1",
+        "eslint": "^8.39.0",
+        "eslint-config-standard-with-typescript": "^34.0.1",
+        "eslint-plugin-import": "^2.26.0",
+        "eslint-plugin-n": "^15.7.0",
+        "eslint-plugin-promise": "^6.1.1",
+        "mocha": "^10.2.0",
+        "openai": "^4.24.1",
+        "sinon": "^15.1.0",
+        "temp": "^0.9.4",
+        "ts-node": "^10.9.1",
+        "ts-node-dev": "^2.0.0",
+        "typedoc": "^0.24.7",
+        "typedoc-plugin-markdown": "^3.15.3",
+        "typescript": "*",
+        "uuid": "^9.0.0"
+      },
+      "optionalDependencies": {
+        "@lancedb/vectordb-darwin-arm64": "0.4.6",
+        "@lancedb/vectordb-darwin-x64": "0.4.6",
+        "@lancedb/vectordb-linux-arm64-gnu": "0.4.6",
+        "@lancedb/vectordb-linux-x64-gnu": "0.4.6",
+        "@lancedb/vectordb-win32-x64-msvc": "0.4.6"
+      }
+    },
+    "../node/node_modules/apache-arrow": {
+      "version": "14.0.2",
+      "license": "Apache-2.0",
+      "dependencies": {
+        "@types/command-line-args": "5.2.0",
+        "@types/command-line-usage": "5.0.2",
+        "@types/node": "20.3.0",
+        "@types/pad-left": "2.1.1",
+        "command-line-args": "5.2.1",
+        "command-line-usage": "7.0.1",
+        "flatbuffers": "23.5.26",
+        "json-bignum": "^0.0.3",
+        "pad-left": "^2.1.0",
+        "tslib": "^2.5.3"
+      },
+      "bin": {
+        "arrow2csv": "bin/arrow2csv.js"
+      }
+    },
+    "node_modules/@types/node": {
+      "version": "20.11.8",
+      "resolved": "https://registry.npmjs.org/@types/node/-/node-20.11.8.tgz",
+      "integrity": "sha512-i7omyekpPTNdv4Jb/Rgqg0RU8YqLcNsI12quKSDkRXNfx7Wxdm6HhK1awT3xTgEkgxPn3bvnSpiEAc7a7Lpyow==",
+      "dev": true,
+      "dependencies": {
+        "undici-types": "~5.26.4"
+      }
+    },
+    "node_modules/apache-arrow": {
+      "resolved": "../node/node_modules/apache-arrow",
+      "link": true
+    },
+    "node_modules/typescript": {
+      "version": "5.3.3",
+      "resolved": "https://registry.npmjs.org/typescript/-/typescript-5.3.3.tgz",
+      "integrity": "sha512-pXWcraxM0uxAS+tN0AG/BF2TyqmHO014Z070UsJ+pFvYuRSq8KH8DmWpnbXe0pEPDHXZV3FcAbJkijJ5oNEnWw==",
+      "dev": true,
+      "bin": {
+        "tsc": "bin/tsc",
+        "tsserver": "bin/tsserver"
+      },
+      "engines": {
+        "node": ">=14.17"
+      }
+    },
+    "node_modules/undici-types": {
+      "version": "5.26.5",
+      "resolved": "https://registry.npmjs.org/undici-types/-/undici-types-5.26.5.tgz",
+      "integrity": "sha512-JlCMO+ehdEIKqlFxk6IfVoAUVmgz7cU7zD/h9XZ0qzeosSHmUJVOzSQvvYSYWXkFXC+IfLKSIffhv0sVZup6pA==",
+      "dev": true
+    },
+    "node_modules/vectordb": {
+      "resolved": "../node",
+      "link": true
+    }
+  }
+}

docs/package.json

@@ -0,0 +1,20 @@
+{
+  "name": "lancedb-docs-test",
+  "version": "1.0.0",
+  "description": "auto-generated tests from doc",
+  "author": "dev@lancedb.com",
+  "license": "Apache 2",
+  "dependencies": {
+    "apache-arrow": "file:../node/node_modules/apache-arrow",
+    "vectordb": "file:../node"
+  },
+  "scripts": {
+    "build": "tsc -b && cd ../node && npm run build-release",
+    "example": "npm run build && node",
+    "test": "npm run build && ls dist/*.js | xargs -n 1 node"
+  },
+  "devDependencies": {
+    "@types/node": "^20.11.8",
+    "typescript": "^5.3.3"
+  }
+}

docs/src/ann_indexes.md

@@ -7,7 +7,7 @@ for brute-force scanning of the entire vector space.
 A vector index is faster but less accurate than exhaustive search (kNN or flat search).
 LanceDB provides many parameters to fine-tune the index's size, the speed of queries, and the accuracy of results.
 
-Currently, LanceDB does *not* automatically create the ANN index.
+Currently, LanceDB does _not_ automatically create the ANN index.
 LanceDB has optimized code for kNN as well. For many use-cases, datasets under 100K vectors won't require index creation at all.
 If you can live with <100ms latency, skipping index creation is a simpler workflow while guaranteeing 100% recall.
 
@@ -17,16 +17,17 @@ In the future we will look to automatically create and configure the ANN index a
 
 Lance can support multiple index types, the most widely used one is `IVF_PQ`.
 
-* `IVF_PQ`: use **Inverted File Index (IVF)** to first divide the dataset into `N` partitions,
-   and then use **Product Quantization** to compress vectors in each partition.
-* `DiskANN` (**Experimental**): organize the vector as a on-disk graph, where the vertices approximately
-   represent the nearest neighbors of each vector.
+- `IVF_PQ`: use **Inverted File Index (IVF)** to first divide the dataset into `N` partitions,
+  and then use **Product Quantization** to compress vectors in each partition.
+- `DiskANN` (**Experimental**): organize the vector as a on-disk graph, where the vertices approximately
+  represent the nearest neighbors of each vector.
 
 ## Creating an IVF_PQ Index
 
 Lance supports `IVF_PQ` index type by default.
 
 === "Python"
+
      Creating indexes is done via the [create_index](https://lancedb.github.io/lancedb/python/#lancedb.table.LanceTable.create_index) method.
 
      ```python
@@ -46,25 +47,20 @@ Lance supports `IVF_PQ` index type by default.
      tbl.create_index(num_partitions=256, num_sub_vectors=96)
      ```
 
-=== "Javascript"
-     ```javascript
-     const vectordb = require('vectordb')
-     const db = await vectordb.connect('data/sample-lancedb')
+=== "Typescript"
 
-     let data = []
-     for (let i = 0; i < 10_000; i++) {
-         data.push({vector: Array(1536).fill(i), id: `${i}`, content: "", longId: `${i}`},)
-     }
-     const table = await db.createTable('my_vectors', data)
-     await table.createIndex({ type: 'ivf_pq', column: 'vector', num_partitions: 256, num_sub_vectors: 96 })
+     ```typescript
+     --8<--- "docs/src/ann_indexes.ts:import"
+
+     --8<-- "docs/src/ann_indexes.ts:ingest"
      ```
 
 - **metric** (default: "L2"): The distance metric to use. By default it uses euclidean distance "`L2`".
-We also support "cosine" and "dot" distance as well.
+  We also support "cosine" and "dot" distance as well.
 - **num_partitions** (default: 256): The number of partitions of the index.
 - **num_sub_vectors** (default: 96): The number of sub-vectors (M) that will be created during Product Quantization (PQ).
-For D dimensional vector, it will be divided into `M` of `D/M` sub-vectors, each of which is presented by
-a single PQ code.
+  For D dimensional vector, it will be divided into `M` of `D/M` sub-vectors, each of which is presented by
+  a single PQ code.
 
 <figure markdown>
   ![IVF PQ](./assets/ivf_pq.png)
@@ -78,7 +74,7 @@ Using GPU for index creation requires [PyTorch>2.0](https://pytorch.org/) being
 
 You can specify the GPU device to train IVF partitions via
 
-- **accelerator**: Specify to ``cuda`` or ``mps`` (on Apple Silicon) to enable GPU training.
+- **accelerator**: Specify to `cuda` or `mps` (on Apple Silicon) to enable GPU training.
 
 === "Linux"
 
@@ -106,10 +102,9 @@ You can specify the GPU device to train IVF partitions via
 
 Trouble shootings:
 
-If you see ``AssertionError: Torch not compiled with CUDA enabled``, you need to [install
+If you see `AssertionError: Torch not compiled with CUDA enabled`, you need to [install
 PyTorch with CUDA support](https://pytorch.org/get-started/locally/).
 
-
 ## Querying an ANN Index
 
 Querying vector indexes is done via the [search](https://lancedb.github.io/lancedb/python/#lancedb.table.LanceTable.search) function.
@@ -127,6 +122,7 @@ There are a couple of parameters that can be used to fine-tune the search:
   Note: refine_factor is only applicable if an ANN index is present. If specified on a table without an ANN index, it is ignored.
 
 === "Python"
+
      ```python
      tbl.search(np.random.random((1536))) \
          .limit(2) \
@@ -134,41 +130,35 @@ There are a couple of parameters that can be used to fine-tune the search:
          .refine_factor(10) \
          .to_pandas()
      ```
-     ```
+
+     ```text
                                               vector       item       _distance
      0  [0.44949695, 0.8444449, 0.06281311, 0.23338133...  item 1141  103.575333
      1  [0.48587373, 0.269207, 0.15095535, 0.65531915,...  item 3953  108.393867
      ```
 
-=== "Javascript"
-     ```javascript
-     const results_1 = await table
-         .search(Array(1536).fill(1.2))
-         .limit(2)
-         .nprobes(20)
-         .refineFactor(10)
-         .execute()
+=== "Typescript"
+
+     ```typescript
+     --8<-- "docs/src/ann_indexes.ts:search1"
      ```
 
 The search will return the data requested in addition to the distance of each item.
 
-
 ### Filtering (where clause)
 
 You can further filter the elements returned by a search using a where clause.
 
 === "Python"
+
      ```python
      tbl.search(np.random.random((1536))).where("item != 'item 1141'").to_pandas()
      ```
 
-=== "Javascript"
+=== "Typescript"
+
      ```javascript
-     const results_2 = await table
-         .search(Array(1536).fill(1.2))
-         .where("id != '1141'")
-         .limit(2)
-         .execute()
+     --8<-- "docs/src/ann_indexes.ts:search2"
      ```
 
 ### Projections (select clause)
@@ -176,23 +166,23 @@ You can further filter the elements returned by a search using a where clause.
 You can select the columns returned by the query using a select clause.
 
 === "Python"
+
      ```python
      tbl.search(np.random.random((1536))).select(["vector"]).to_pandas()
      ```
-     ```
-        vector                                             _distance
+
+
+     ```text
+                                                   vector _distance
      0  [0.30928212, 0.022668175, 0.1756372, 0.4911822...  93.971092
      1  [0.2525465, 0.01723831, 0.261568, 0.002007689,...  95.173485
      ...
      ```
 
-=== "Javascript"
-     ```javascript
-     const results_3 = await table
-         .search(Array(1536).fill(1.2))
-         .select(["id"])
-         .limit(2)
-         .execute()
+=== "Typescript"
+
+     ```typescript
+     --8<-- "docs/src/ann_indexes.ts:search3"
      ```
 
 ## FAQ
@@ -221,4 +211,4 @@ On `SIFT-1M` dataset, our benchmark shows that keeping each partition 1K-4K rows
 `num_sub_vectors` specifies how many Product Quantization (PQ) short codes to generate on each vector. Because
 PQ is a lossy compression of the original vector, a higher `num_sub_vectors` usually results in
 less space distortion, and thus yields better accuracy. However, a higher `num_sub_vectors` also causes heavier I/O and
-more PQ computation, and thus, higher latency. `dimension / num_sub_vectors` should be a multiple of 8 for optimum SIMD efficiency.
+more PQ computation, and thus, higher latency. `dimension / num_sub_vectors` should be a multiple of 8 for optimum SIMD efficiency.

docs/src/ann_indexes.ts

@@ -0,0 +1,53 @@
+// --8<-- [start:import]
+import * as vectordb from "vectordb";
+// --8<-- [end:import]
+
+(async () => {
+  // --8<-- [start:ingest]
+  const db = await vectordb.connect("data/sample-lancedb");
+
+  let data = [];
+  for (let i = 0; i < 10_000; i++) {
+    data.push({
+      vector: Array(1536).fill(i),
+      id: `${i}`,
+      content: "",
+      longId: `${i}`,
+    });
+  }
+  const table = await db.createTable("my_vectors", data);
+  await table.createIndex({
+    type: "ivf_pq",
+    column: "vector",
+    num_partitions: 16,
+    num_sub_vectors: 48,
+  });
+  // --8<-- [end:ingest]
+
+  // --8<-- [start:search1]
+  const results_1 = await table
+    .search(Array(1536).fill(1.2))
+    .limit(2)
+    .nprobes(20)
+    .refineFactor(10)
+    .execute();
+  // --8<-- [end:search1]
+
+  // --8<-- [start:search2]
+  const results_2 = await table
+    .search(Array(1536).fill(1.2))
+    .where("id != '1141'")
+    .limit(2)
+    .execute();
+  // --8<-- [end:search2]
+
+  // --8<-- [start:search3]
+  const results_3 = await table
+    .search(Array(1536).fill(1.2))
+    .select(["id"])
+    .limit(2)
+    .execute();
+  // --8<-- [end:search3]
+
+  console.log("Ann indexes: done");
+})();

docs/src/basic.md

@@ -11,43 +11,78 @@
 ## Installation
 
 === "Python"
+
       ```shell
       pip install lancedb
       ```
 
-=== "Javascript"
+=== "Typescript"
+
       ```shell
       npm install vectordb
       ```
 
+=== "Rust"
+
+    !!! warning "Rust SDK is experimental, might introduce breaking changes in the near future"
+
+    ```shell
+    cargo add vectordb
+    ```
+
+    !!! info "To use the vectordb create, you first need to install protobuf."
+
+    === "macOS"
+
+        ```shell
+        brew install protobuf
+        ```
+
+    === "Ubuntu/Debian"
+
+        ```shell
+        sudo apt install -y protobuf-compiler libssl-dev
+        ```
+
+    !!! info "Please also make sure you're using the same version of Arrow as in the [vectordb crate](https://github.com/lancedb/lancedb/blob/main/Cargo.toml)"
+
 ## How to connect to a database
 
 === "Python"
+
       ```python
       import lancedb
       uri = "data/sample-lancedb"
       db = lancedb.connect(uri)
       ```
 
-      LanceDB will create the directory if it doesn't exist (including parent directories).
+=== "Typescript"
 
-      If you need a reminder of the uri, use the `db.uri` property.
+    ```typescript
+    --8<-- "docs/src/basic_legacy.ts:import"
 
-=== "Javascript"
-      ```javascript
-      const lancedb = require("vectordb");
+    --8<-- "docs/src/basic_legacy.ts:open_db"
+    ```
 
-      const uri = "data/sample-lancedb";
-      const db = await lancedb.connect(uri);
-      ```
-      
-      LanceDB will create the directory if it doesn't exist (including parent directories).
+=== "Rust"
 
-      If you need a reminder of the uri, you can call `db.uri()`.
+    ```rust
+    #[tokio::main]
+    async fn main() -> Result<()> {
+        --8<-- "rust/vectordb/examples/simple.rs:connect"
+    }
+    ```
+
+    !!! info "See [examples/simple.rs](https://github.com/lancedb/lancedb/tree/main/rust/vectordb/examples/simple.rs) for a full working example."
+
+LanceDB will create the directory if it doesn't exist (including parent directories).
+
+If you need a reminder of the uri, you can call `db.uri()`.
 
 ## How to create a table
 
 === "Python"
+
     ```python
     tbl = db.create_table("my_table",
                     data=[{"vector": [3.1, 4.1], "item": "foo", "price": 10.0},
@@ -59,6 +94,7 @@
     to the `create_table` method.
 
     You can also pass in a pandas DataFrame directly:
+
     ```python
     import pandas as pd
     df = pd.DataFrame([{"vector": [3.1, 4.1], "item": "foo", "price": 10.0},
@@ -66,19 +102,26 @@
     tbl = db.create_table("table_from_df", data=df)
     ```
 
-=== "Javascript"
-    ```javascript
-    const tb = await db.createTable(
-        "myTable",
-        [{"vector": [3.1, 4.1], "item": "foo", "price": 10.0},
-         {"vector": [5.9, 26.5], "item": "bar", "price": 20.0}]
-    )
+=== "Typescript"
+
+    ```typescript
+    --8<-- "docs/src/basic_legacy.ts:create_table"
     ```
 
     If the table already exists, LanceDB will raise an error by default.
     If you want to overwrite the table, you can pass in `mode="overwrite"`
     to the `createTable` function.
 
+=== "Rust"
+
+    ```rust
+    use arrow_schema::{DataType, Schema, Field};
+    use arrow_array::{RecordBatch, RecordBatchIterator};
+
+    --8<-- "rust/vectordb/examples/simple.rs:create_table"
+    ```
+
+    If the table already exists, LanceDB will raise an error by default.
 
 !!! info "Under the hood, LanceDB is converting the input data into an Apache Arrow table and persisting it to disk in [Lance format](https://www.github.com/lancedb/lance)."
 
@@ -88,76 +131,145 @@ Sometimes you may not have the data to insert into the table at creation time.
 In this case, you can create an empty table and specify the schema.
 
 === "Python"
+
       ```python
       import pyarrow as pa
       schema = pa.schema([pa.field("vector", pa.list_(pa.float32(), list_size=2))])
       tbl = db.create_table("empty_table", schema=schema)
       ```
 
+=== "Typescript"
+
+    ```typescript
+    --8<-- "docs/src/basic_legacy.ts:create_empty_table"
+    ```
+
+=== "Rust"
+
+    ```rust
+    --8<-- "rust/vectordb/examples/simple.rs:create_empty_table"
+    ```
+
 ## How to open an existing table
 
 Once created, you can open a table using the following code:
 
 === "Python"
-      ```python
-      tbl = db.open_table("my_table")
-      ```
 
-      If you forget the name of your table, you can always get a listing of all table names:
+    ```python
+    tbl = db.open_table("my_table")
+    ```
 
-      ```python
-      print(db.table_names())
-      ```
+=== "Typescript"
+
+    ```typescript
+    const tbl = await db.openTable("myTable");
+    ```
+
+=== "Rust"
+
+    ```rust
+    --8<-- "rust/vectordb/examples/simple.rs:open_with_existing_file"
+    ```
+
+If you forget the name of your table, you can always get a listing of all table names:
+
+=== "Python"
+
+    ```python
+    print(db.table_names())
+    ```
 
 === "Javascript"
-      ```javascript
-      const tbl = await db.openTable("myTable");
-      ```
 
-      If you forget the name of your table, you can always get a listing of all table names:
+    ```javascript
+    console.log(await db.tableNames());
+    ```
 
-      ```javascript
-      console.log(await db.tableNames());
-      ```
+=== "Rust"
+
+    ```rust
+    --8<-- "rust/vectordb/examples/simple.rs:list_names"
+    ```
 
 ## How to add data to a table
 
 After a table has been created, you can always add more data to it using
 
 === "Python"
-      ```python
 
-      # Option 1: Add a list of dicts to a table
-      data = [{"vector": [1.3, 1.4], "item": "fizz", "price": 100.0},
-              {"vector": [9.5, 56.2], "item": "buzz", "price": 200.0}]
-      tbl.add(data)
+    ```python
 
-      # Option 2: Add a pandas DataFrame to a table
-      df = pd.DataFrame(data)
-      tbl.add(data)
-      ```
+    # Option 1: Add a list of dicts to a table
+    data = [{"vector": [1.3, 1.4], "item": "fizz", "price": 100.0},
+            {"vector": [9.5, 56.2], "item": "buzz", "price": 200.0}]
+    tbl.add(data)
 
-=== "Javascript"
-      ```javascript
-      await tbl.add([{vector: [1.3, 1.4], item: "fizz", price: 100.0},
-                     {vector: [9.5, 56.2], item: "buzz", price: 200.0}])
-      ```
+    # Option 2: Add a pandas DataFrame to a table
+    df = pd.DataFrame(data)
+    tbl.add(data)
+    ```
+
+=== "Typescript"
+
+    ```typescript
+    --8<-- "docs/src/basic_legacy.ts:add"
+    ```
+
+=== "Rust"
+
+    ```rust
+    --8<-- "rust/vectordb/examples/simple.rs:add"
+    ```
 
 ## How to search for (approximate) nearest neighbors
 
 Once you've embedded the query, you can find its nearest neighbors using the following code:
 
 === "Python"
-      ```python
-      tbl.search([100, 100]).limit(2).to_pandas()
-      ```
 
-      This returns a pandas DataFrame with the results.
+    ```python
+    tbl.search([100, 100]).limit(2).to_pandas()
+    ```
 
-=== "Javascript"
-      ```javascript
-      const query = await tbl.search([100, 100]).limit(2).execute();
-      ```
+    This returns a pandas DataFrame with the results.
+
+=== "Typescript"
+
+    ```typescript
+    --8<-- "docs/src/basic_legacy.ts:search"
+    ```
+
+=== "Rust"
+
+    ```rust
+    use futures::TryStreamExt;
+
+    --8<-- "rust/vectordb/examples/simple.rs:search"
+    ```
+
+By default, LanceDB runs a brute-force scan over dataset to find the K nearest neighbours (KNN).
+For tables with more than 50K vectors, creating an ANN index is recommended to speed up search performance.
+
+=== "Python"
+
+    ```py
+    tbl.create_index()
+    ```
+
+=== "Typescript"
+
+    ```{.typescript .ignore}
+    --8<-- "docs/src/basic_legacy.ts:create_index"
+    ```
+
+=== "Rust"
+
+    ```rust
+     --8<-- "rust/vectordb/examples/simple.rs:create_index"
+    ```
+
+Check [Approximate Nearest Neighbor (ANN) Indexes](/ann_indices.md) section for more details.
 
 ## How to delete rows from a table
 
@@ -166,20 +278,27 @@ which rows to delete, provide a filter that matches on the metadata columns.
 This can delete any number of rows that match the filter.
 
 === "Python"
-      ```python
-      tbl.delete('item = "fizz"')
-      ```
 
-=== "Javascript"
-      ```javascript
-      await tbl.delete('item = "fizz"')
-      ```
+    ```python
+    tbl.delete('item = "fizz"')
+    ```
+
+=== "Typescript"
+
+    ```typescript
+    --8<-- "docs/src/basic_legacy.ts:delete"
+    ```
+
+=== "Rust"
+
+    ```rust
+    --8<-- "rust/vectordb/examples/simple.rs:delete"
+    ```
 
 The deletion predicate is a SQL expression that supports the same expressions
 as the `where()` clause on a search. They can be as simple or complex as needed.
 To see what expressions are supported, see the [SQL filters](sql.md) section.
 
-
 === "Python"
 
       Read more: [lancedb.table.Table.delete][]
@@ -193,6 +312,7 @@ To see what expressions are supported, see the [SQL filters](sql.md) section.
 Use the `drop_table()` method on the database to remove a table.
 
 === "Python"
+
       ```python
       db.drop_table("my_table")
       ```
@@ -201,13 +321,20 @@ Use the `drop_table()` method on the database to remove a table.
       By default, if the table does not exist an exception is raised. To suppress this,
       you can pass in `ignore_missing=True`.
 
-=== "JavaScript"
-      ```javascript
-      await db.dropTable('myTable')
+=== "Typescript"
+
+      ```typescript
+      --8<-- "docs/src/basic_legacy.ts:drop_table"
       ```
 
       This permanently removes the table and is not recoverable, unlike deleting rows.
-      If the table does not exist an exception is raised. 
+      If the table does not exist an exception is raised.
+
+=== "Rust"
+
+    ```rust
+    --8<-- "rust/vectordb/examples/simple.rs:drop_table"
+    ```
 
 !!! note "Bundling `vectordb` apps with Webpack"
 

docs/src/basic_legacy.ts

@@ -0,0 +1,92 @@
+// --8<-- [start:import]
+import * as lancedb from "vectordb";
+import { Schema, Field, Float32, FixedSizeList, Int32, Float16 } from "apache-arrow";
+// --8<-- [end:import]
+import * as fs from "fs";
+import { Table as ArrowTable, Utf8 } from "apache-arrow";
+
+const example = async () => {
+  fs.rmSync("data/sample-lancedb", { recursive: true, force: true });
+  // --8<-- [start:open_db]
+  const lancedb = require("vectordb");
+  const uri = "data/sample-lancedb";
+  const db = await lancedb.connect(uri);
+  // --8<-- [end:open_db]
+
+  // --8<-- [start:create_table]
+  const tbl = await db.createTable(
+    "myTable",
+    [
+      { vector: [3.1, 4.1], item: "foo", price: 10.0 },
+      { vector: [5.9, 26.5], item: "bar", price: 20.0 },
+    ],
+    { writeMode: lancedb.WriteMode.Overwrite }
+  );
+  // --8<-- [end:create_table]
+
+  // --8<-- [start:add]
+  const newData = Array.from({ length: 500 }, (_, i) => ({
+    vector: [i, i + 1],
+    item: "fizz",
+    price: i * 0.1,
+  }));
+  await tbl.add(newData);
+  // --8<-- [end:add]
+
+  // --8<-- [start:create_index]
+  await tbl.createIndex({
+    type: "ivf_pq",
+    num_partitions: 2,
+    num_sub_vectors: 2,
+  });
+  // --8<-- [end:create_index]
+
+  // --8<-- [start:create_empty_table]
+  const schema = new Schema([
+    new Field("id", new Int32()),
+    new Field("name", new Utf8()),
+  ]);
+  const empty_tbl = await db.createTable({ name: "empty_table", schema });
+  // --8<-- [end:create_empty_table]
+
+  // --8<-- [start:create_f16_table]
+  const dim = 16
+  const total = 10
+  const f16_schema = new Schema([
+      new Field('id', new Int32()),
+      new Field(
+        'vector',
+        new FixedSizeList(dim, new Field('item', new Float16(), true)),
+        false
+      )
+    ])
+  const data = lancedb.makeArrowTable(
+      Array.from(Array(total), (_, i) => ({
+        id: i,
+        vector: Array.from(Array(dim), Math.random)
+      })),
+      { f16_schema }
+    )
+  const table = await db.createTable('f16_tbl', data)
+  // --8<-- [end:create_f16_table]
+
+  // --8<-- [start:search]
+  const query = await tbl.search([100, 100]).limit(2).execute();
+  // --8<-- [end:search]
+  console.log(query);
+
+  // --8<-- [start:delete]
+  await tbl.delete('item = "fizz"');
+  // --8<-- [end:delete]
+
+  // --8<-- [start:drop_table]
+  await db.dropTable("myTable");
+  // --8<-- [end:drop_table]
+};
+
+async function main() {
+  await example();
+  console.log("Basic example: done");
+}
+
+main();

docs/src/embeddings/api.md

@@ -17,6 +17,7 @@ Let's implement `SentenceTransformerEmbeddings` class. All you need to do is imp
 
 ```python
 from lancedb.embeddings import register
+from lancedb.util import attempt_import_or_raise
 
 @register("sentence-transformers")
 class SentenceTransformerEmbeddings(TextEmbeddingFunction):
@@ -81,7 +82,7 @@ class OpenClipEmbeddings(EmbeddingFunction):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        open_clip = self.safe_import("open_clip", "open-clip") # EmbeddingFunction util to import external libs and raise if not found
+        open_clip = attempt_import_or_raise("open_clip", "open-clip") # EmbeddingFunction util to import external libs and raise if not found
         model, _, preprocess = open_clip.create_model_and_transforms(
             self.name, pretrained=self.pretrained
         )
@@ -109,14 +110,14 @@ class OpenClipEmbeddings(EmbeddingFunction):
         if isinstance(query, str):
             return [self.generate_text_embeddings(query)]
         else:
-            PIL = self.safe_import("PIL", "pillow")
+            PIL = attempt_import_or_raise("PIL", "pillow")
             if isinstance(query, PIL.Image.Image):
                 return [self.generate_image_embedding(query)]
             else:
                 raise TypeError("OpenClip supports str or PIL Image as query")
 
     def generate_text_embeddings(self, text: str) -> np.ndarray:
-        torch = self.safe_import("torch")
+        torch = attempt_import_or_raise("torch")
         text = self.sanitize_input(text)
         text = self._tokenizer(text)
         text.to(self.device)
@@ -175,7 +176,7 @@ class OpenClipEmbeddings(EmbeddingFunction):
             The image to embed. If the image is a str, it is treated as a uri.
             If the image is bytes, it is treated as the raw image bytes.
         """
-        torch = self.safe_import("torch")
+        torch = attempt_import_or_raise("torch")
         # TODO handle retry and errors for https
         image = self._to_pil(image)
         image = self._preprocess(image).unsqueeze(0)
@@ -183,7 +184,7 @@ class OpenClipEmbeddings(EmbeddingFunction):
             return self._encode_and_normalize_image(image)
 
     def _to_pil(self, image: Union[str, bytes]):
-        PIL = self.safe_import("PIL", "pillow")
+        PIL = attempt_import_or_raise("PIL", "pillow")
         if isinstance(image, bytes):
             return PIL.Image.open(io.BytesIO(image))
         if isinstance(image, PIL.Image.Image):

docs/src/embeddings/default_embedding_functions.md

@@ -9,6 +9,9 @@ Contains the text embedding functions registered by default.
 ### Sentence transformers
 Allows you to set parameters when registering a `sentence-transformers` object.
 
+!!! info
+    Sentence transformer embeddings are normalized by default. It is recommended to use normalized embeddings for similarity search.
+
 | Parameter | Type | Default Value | Description |
 |---|---|---|---|
 | `name` | `str` | `all-MiniLM-L6-v2` | The name of the model |
@@ -119,7 +122,7 @@ texts = [{"text": "Capitalism has been dominant in the Western world since the e
 tbl.add(texts)
 ```
 
-## Gemini Embedding Function
+### Gemini Embeddings
 With Google's Gemini, you can represent text (words, sentences, and blocks of text) in a vectorized form, making it easier to compare and contrast embeddings. For example, two texts that share a similar subject matter or sentiment should have similar embeddings, which can be identified through mathematical comparison techniques such as cosine similarity. For more on how and why you should use embeddings, refer to the Embeddings guide.
 The Gemini Embedding Model API supports various task types:
 
@@ -155,6 +158,51 @@ tbl.add(df)
 rs = tbl.search("hello").limit(1).to_pandas()
 ```
 
+### AWS Bedrock Text Embedding Functions
+AWS Bedrock supports multiple base models for generating text embeddings. You need to setup the AWS credentials to use this embedding function.
+You can do so by using `awscli` and also add your session_token:
+```shell
+aws configure
+aws configure set aws_session_token "<your_session_token>"
+```
+to ensure that the credentials are set up correctly, you can run the following command:
+```shell
+aws sts get-caller-identity
+```
+
+Supported Embedding modelIDs are:
+* `amazon.titan-embed-text-v1`
+* `cohere.embed-english-v3`
+* `cohere.embed-multilingual-v3`
+
+Supported paramters (to be passed in `create` method) are:
+| Parameter | Type | Default Value | Description |
+|---|---|---|---|
+| **name** | str | "amazon.titan-embed-text-v1" | The model ID of the bedrock model to use. Supported base models for Text Embeddings: amazon.titan-embed-text-v1, cohere.embed-english-v3, cohere.embed-multilingual-v3 |
+| **region** | str | "us-east-1" | Optional name of the AWS Region in which the service should be called (e.g., "us-east-1"). |
+| **profile_name** | str | None | Optional name of the AWS profile to use for calling the Bedrock service. If not specified, the default profile will be used. |
+| **assumed_role** | str | None | Optional ARN of an AWS IAM role to assume for calling the Bedrock service. If not specified, the current active credentials will be used. |
+| **role_session_name** | str | "lancedb-embeddings" | Optional name of the AWS IAM role session to use for calling the Bedrock service. If not specified, a "lancedb-embeddings" name will be used. |
+| **runtime** | bool | True | Optional choice of getting different client to perform operations with the Amazon Bedrock service. |
+| **max_retries** | int | 7 | Optional number of retries to perform when a request fails. |
+
+Usage Example:
+
+```python
+model = get_registry().get("bedrock-text").create()
+
+class TextModel(LanceModel):
+    text: str = model.SourceField()
+    vector: Vector(model.ndims()) = model.VectorField()
+
+df = pd.DataFrame({"text": ["hello world", "goodbye world"]})
+db = lancedb.connect("tmp_path")
+tbl = db.create_table("test", schema=TextModel, mode="overwrite")
+
+tbl.add(df)
+rs = tbl.search("hello").limit(1).to_pandas()
+```
+
 ## Multi-modal embedding functions
 Multi-modal embedding functions allow you to query your table using both images and text.
 

docs/src/examples/modal_langchain.py

@@ -79,7 +79,10 @@ def qanda_langchain(query):
     download_docs()
     docs = store_docs()
 
-    text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200,)
+    text_splitter = RecursiveCharacterTextSplitter(
+        chunk_size=1000,
+        chunk_overlap=200,
+    )
     documents = text_splitter.split_documents(docs)
     embeddings = OpenAIEmbeddings()
 

docs/src/extra_js/init_ask_ai_widget.js

@@ -0,0 +1,11 @@
+document.addEventListener("DOMContentLoaded", function () {
+    var script = document.createElement("script");
+    script.src = "https://widget.kapa.ai/kapa-widget.bundle.js";
+    script.setAttribute("data-website-id", "c5881fae-cec0-490b-b45e-d83d131d4f25");
+    script.setAttribute("data-project-name", "LanceDB");
+    script.setAttribute("data-project-color", "#000000");
+    script.setAttribute("data-project-logo", "https://avatars.githubusercontent.com/u/108903835?s=200&v=4");
+  script.setAttribute("data-modal-example-questions","Help me create an IVF_PQ index,How do I do an exhaustive search?,How do I create a LanceDB table?,Can I use my own embedding function?");
+    script.async = true;
+    document.head.appendChild(script);
+  });

docs/src/faq.md

@@ -69,3 +69,19 @@ MinIO supports an S3 compatible API. In order to connect to a MinIO instance, yo
 - Set the envvar `AWS_ENDPOINT` to the URL of your MinIO API
 - Set the envvars `AWS_ACCESS_KEY_ID` and `AWS_SECRET_ACCESS_KEY` with your MinIO credential
 - Call `lancedb.connect("s3://minio_bucket_name")`
+
+### Where can I find benchmarks for LanceDB?
+
+Refer to this [post](https://blog.lancedb.com/benchmarking-lancedb-92b01032874a) for recent benchmarks.
+
+### How much data can LanceDB practically manage without effecting performance?
+
+We target good performance on ~10-50 billion rows and ~10-30 TB of data.
+
+### Does LanceDB support concurrent operations?
+
+LanceDB can handle concurrent reads very well, and can scale horizontally. The main constraint is how well the [storage layer](https://lancedb.github.io/lancedb/concepts/storage/) you've chosen scales. For writes, we support concurrent writing, though too many concurrent writers can lead to failing writes as there is a limited number of times a writer retries a commit
+
+!!! info "Multiprocessing with LanceDB"
+
+    For multiprocessing you should probably not use ```fork``` as lance is multi-threaded internally and ```fork``` and multi-thread do not work well.[Refer to this discussion](https://discuss.python.org/t/concerns-regarding-deprecation-of-fork-with-alive-threads/33555)

docs/src/guides/storage.md

@@ -68,6 +68,82 @@ Alternatively, if you are using AWS SSO, you can use the `AWS_PROFILE` and `AWS_
 
 You can see a full list of environment variables [here](https://docs.rs/object_store/latest/object_store/aws/struct.AmazonS3Builder.html#method.from_env).
 
+!!! tip "Automatic cleanup for failed writes"
+
+    LanceDB uses [multi-part uploads](https://docs.aws.amazon.com/AmazonS3/latest/userguide/mpuoverview.html) when writing data to S3 in order to maximize write speed. LanceDB will abort these uploads when it shuts down gracefully, such as when cancelled by keyboard interrupt. However, in the rare case that LanceDB crashes, it is possible that some data will be left lingering in your account. To cleanup this data, we recommend (as AWS themselves do) that you setup a lifecycle rule to delete in-progress uploads after 7 days. See the AWS guide:
+
+    **[Configuring a bucket lifecycle configuration to delete incomplete multipart uploads](https://docs.aws.amazon.com/AmazonS3/latest/userguide/mpu-abort-incomplete-mpu-lifecycle-config.html)**
+
+#### AWS IAM Permissions
+
+If a bucket is private, then an IAM policy must be specified to allow access to it. For many development scenarios, using broad permissions such as a PowerUser account is more than sufficient for working with LanceDB. However, in many production scenarios, you may wish to have as narrow as possible permissions.
+
+For **read and write access**, LanceDB will need a policy such as:
+
+```json
+{
+    "Version": "2012-10-17",
+    "Statement": [
+        {
+            "Effect": "Allow",
+            "Action": [
+              "s3:PutObject",
+              "s3:GetObject",
+              "s3:DeleteObject",
+            ],
+            "Resource": "arn:aws:s3:::<bucket>/<prefix>/*"
+        },
+        {
+            "Effect": "Allow",
+            "Action": [
+                "s3:ListBucket",
+                "s3:GetBucketLocation"
+            ],
+            "Resource": "arn:aws:s3:::<bucket>",
+            "Condition": {
+                "StringLike": {
+                    "s3:prefix": [
+                        "<prefix>/*"
+                    ]
+                }
+            }
+        }
+    ]
+}
+```
+
+For **read-only access**, LanceDB will need a policy such as:
+
+```json
+{
+    "Version": "2012-10-17",
+    "Statement": [
+        {
+            "Effect": "Allow",
+            "Action": [
+              "s3:GetObject",
+            ],
+            "Resource": "arn:aws:s3:::<bucket>/<prefix>/*"
+        },
+        {
+            "Effect": "Allow",
+            "Action": [
+                "s3:ListBucket",
+                "s3:GetBucketLocation"
+            ],
+            "Resource": "arn:aws:s3:::<bucket>",
+            "Condition": {
+                "StringLike": {
+                    "s3:prefix": [
+                        "<prefix>/*"
+                    ]
+                }
+            }
+        }
+    ]
+}
+```
+
 #### S3-compatible stores
 
 LanceDB can also connect to S3-compatible stores, such as MinIO. To do so, you must specify two environment variables: `AWS_ENDPOINT` and `AWS_DEFAULT_REGION`. `AWS_ENDPOINT` should be the URL of the S3-compatible store, and `AWS_DEFAULT_REGION` should be the region to use.

docs/src/guides/tables.md

@@ -16,9 +16,22 @@ This guide will show how to create tables, insert data into them, and update the
     db = lancedb.connect("./.lancedb")
     ```
 
+=== "Javascript"
+
+    Initialize a VectorDB connection and create a table using one of the many methods listed below.
+
+    ```javascript
+    const lancedb = require("vectordb");
+
+    const uri = "data/sample-lancedb";
+    const db = await lancedb.connect(uri);
+    ```
+
     LanceDB allows ingesting data from various sources - `dict`, `list[dict]`, `pd.DataFrame`, `pa.Table` or a `Iterator[pa.RecordBatch]`. Let's take a look at some of the these.
 
-    ### From list of tuples or dictionaries
+### From list of tuples or dictionaries
+
+=== "Python"
 
     ```python
     import lancedb
@@ -32,7 +45,6 @@ This guide will show how to create tables, insert data into them, and update the
 
     db["my_table"].head()
     ```
-
     !!! info "Note"
         If the table already exists, LanceDB will raise an error by default. 
 
@@ -51,6 +63,27 @@ This guide will show how to create tables, insert data into them, and update the
         db.create_table("name", data, mode="overwrite")
         ```
 
+=== "Javascript"
+    You can create a LanceDB table in JavaScript using an array of JSON records as follows.
+
+    ```javascript
+    const tb = await db.createTable("my_table", [{
+        "vector": [3.1, 4.1],
+        "item": "foo",
+        "price": 10.0
+    }, {
+        "vector": [5.9, 26.5],
+        "item": "bar",
+        "price": 20.0
+    }]);
+    ```
+    !!! info "Note"
+    If the table already exists, LanceDB will raise an error by default. If you want to overwrite the table, you need to specify the `WriteMode` in the createTable function.
+
+    ```javascript
+    const table = await con.createTable(tableName, data, { writeMode: WriteMode.Overwrite })
+    ```
+
     ### From a Pandas DataFrame
 
     ```python
@@ -67,7 +100,9 @@ This guide will show how to create tables, insert data into them, and update the
     db["my_table"].head()
     ```
     !!! info "Note"
-        Data is converted to Arrow before being written to disk. For maximum control over how data is saved, either provide the PyArrow schema to convert to or else provide a PyArrow Table directly.
+    Data is converted to Arrow before being written to disk. For maximum control over how data is saved, either provide the PyArrow schema to convert to or else provide a PyArrow Table directly.
+
+    The **`vector`** column needs to be a [Vector](../python/pydantic.md#vector-field) (defined as [pyarrow.FixedSizeList](https://arrow.apache.org/docs/python/generated/pyarrow.list_.html)) type. 
 
     ```python
     custom_schema = pa.schema([
@@ -79,7 +114,7 @@ This guide will show how to create tables, insert data into them, and update the
     table = db.create_table("my_table", data, schema=custom_schema)
     ```
 
-    ### From a Polars DataFrame
+### From a Polars DataFrame
 
     LanceDB supports [Polars](https://pola.rs/), a modern, fast DataFrame library
     written in Rust. Just like in Pandas, the Polars integration is enabled by PyArrow
@@ -97,26 +132,44 @@ This guide will show how to create tables, insert data into them, and update the
     table = db.create_table("pl_table", data=data)
     ```
 
-    ### From PyArrow Tables
-    You can also create LanceDB tables directly from PyArrow tables
+### From an Arrow Table
+=== "Python"
+    You can also create LanceDB tables directly from Arrow tables. 
+    LanceDB supports float16 data type!
 
     ```python
-    table = pa.Table.from_arrays(
-            [
-                pa.array([[3.1, 4.1, 5.1, 6.1], [5.9, 26.5, 4.7, 32.8]],
-                        pa.list_(pa.float32(), 4)),
-                pa.array(["foo", "bar"]),
-                pa.array([10.0, 20.0]),
-            ],
-            ["vector", "item", "price"],
-        )
+    import pyarrows as pa
+    import numpy as np
+    
+    dim = 16
+    total = 2
+    schema = pa.schema(
+        [
+            pa.field("vector", pa.list_(pa.float16(), dim)),
+            pa.field("text", pa.string())
+        ]
+    )
+    data = pa.Table.from_arrays(
+        [
+            pa.array([np.random.randn(dim).astype(np.float16) for _ in range(total)],
+                    pa.list_(pa.float16(), dim)),
+            pa.array(["foo", "bar"])
+        ],
+        ["vector", "text"],
+    )
+    tbl = db.create_table("f16_tbl", data, schema=schema)
+    ```
 
-    db = lancedb.connect("db")
+=== "Javascript"
+    You can also create LanceDB tables directly from Arrow tables. 
+    LanceDB supports Float16 data type!
 
-    tbl = db.create_table("my_table", table)
+    ```javascript
+    --8<-- "docs/src/basic_legacy.ts:create_f16_table"
     ```
 
     ### From Pydantic Models
+
     When you create an empty table without data, you must specify the table schema.
     LanceDB supports creating tables by specifying a PyArrow schema or a specialized
     Pydantic model called `LanceModel`.
@@ -261,37 +314,6 @@ This guide will show how to create tables, insert data into them, and update the
 
     You can also use iterators of other types like Pandas DataFrame or Pylists directly in the above example.
 
-=== "JavaScript"
-    Initialize a VectorDB connection and create a table using one of the many methods listed below.
-
-    ```javascript
-    const lancedb = require("vectordb");
-
-    const uri = "data/sample-lancedb";
-    const db = await lancedb.connect(uri);
-    ```
-
-    You can create a LanceDB table in JavaScript using an array of JSON records as follows.
-
-    ```javascript
-    const tb = await db.createTable("my_table", [{
-        "vector": [3.1, 4.1],
-        "item": "foo",
-        "price": 10.0
-    }, {
-        "vector": [5.9, 26.5],
-        "item": "bar",
-        "price": 20.0
-    }]);
-    ```
-
-    !!! info "Note"
-    If the table already exists, LanceDB will raise an error by default. If you want to overwrite the table, you need to specify the `WriteMode` in the createTable function.
-
-    ```javascript
-    const table = await con.createTable(tableName, data, { writeMode: WriteMode.Overwrite })
-    ```
-
 ## Open existing tables
 
 === "Python"

docs/src/hybrid_search/hybrid_search.md

@@ -0,0 +1,242 @@
+# Hybrid Search
+
+LanceDB supports both semantic and keyword-based search (also termed full-text search, or FTS). In real world applications, it is often useful to combine these two approaches to get the best best results. For example, you may want to search for a document that is semantically similar to a query document, but also contains a specific keyword. This is an example of *hybrid search*, a search algorithm that combines multiple search techniques.
+
+## Hybrid search in LanceDB
+You can perform hybrid search in LanceDB by combining the results of semantic and full-text search via a reranking algorithm of your choice. LanceDB provides multiple rerankers out of the box. However, you can always write a custom reranker if your use case need more sophisticated logic .
+
+```python
+import os
+
+import lancedb
+import openai
+from lancedb.embeddings import get_registry
+from lancedb.pydantic import LanceModel, Vector
+
+db = lancedb.connect("~/.lancedb")
+
+# Ingest embedding function in LanceDB table
+# Configuring the environment variable OPENAI_API_KEY
+if "OPENAI_API_KEY" not in os.environ:
+# OR set the key here as a variable
+    openai.api_key = "sk-..."
+embeddings = get_registry().get("openai").create()
+
+class Documents(LanceModel):
+    vector: Vector(embeddings.ndims()) = embeddings.VectorField()
+    text: str = embeddings.SourceField()
+
+table = db.create_table("documents", schema=Documents)
+
+data = [
+    { "text": "rebel spaceships striking from a hidden base"},
+    { "text": "have won their first victory against the evil Galactic Empire"},
+    { "text": "during the battle rebel spies managed to steal secret plans"},
+    { "text": "to the Empire's ultimate weapon the Death Star"}
+]
+
+# ingest docs with auto-vectorization
+table.add(data)
+
+# Create a fts index before the hybrid search
+table.create_fts_index("text")
+# hybrid search with default re-ranker
+results = table.search("flower moon", query_type="hybrid").to_pandas()
+```
+
+By default, LanceDB uses `LinearCombinationReranker(weight=0.7)` to combine and rerank the results of semantic and full-text search. You can customize the hyperparameters as needed or write your own custom reranker. Here's how you can use any of the available rerankers:
+
+
+### `rerank()` arguments
+* `normalize`: `str`, default `"score"`:
+    The method to normalize the scores. Can be "rank" or "score". If "rank", the scores are converted to ranks and then normalized. If "score", the scores are normalized directly.
+* `reranker`: `Reranker`, default `LinearCombinationReranker(weight=0.7)`.
+    The reranker to use. If not specified, the default reranker is used.
+
+
+## Available Rerankers
+LanceDB provides a number of re-rankers out of the box. You can use any of these re-rankers by passing them to the `rerank()` method. Here's a list of available re-rankers:
+
+### Linear Combination Reranker
+This is the default re-ranker used by LanceDB. It combines the results of semantic and full-text search using a linear combination of the scores. The weights for the linear combination can be specified. It defaults to 0.7, i.e, 70% weight for semantic search and 30% weight for full-text search.
+
+
+```python
+from lancedb.rerankers import LinearCombinationReranker
+
+reranker = LinearCombinationReranker(weight=0.3) # Use 0.3 as the weight for vector search
+
+results = table.search("rebel", query_type="hybrid").rerank(reranker=reranker).to_pandas()
+```
+
+### Arguments
+----------------
+* `weight`: `float`, default `0.7`:
+    The weight to use for the semantic search score. The weight for the full-text search score is `1 - weights`.
+* `fill`: `float`, default `1.0`:
+        The score to give to results that are only in one of the two result sets.This is treated as penalty, so a higher value means a lower score.
+        TODO: We should just hardcode this-- its pretty confusing as we invert scores to calculate final score
+* `return_score` : str, default `"relevance"`
+        options are "relevance" or "all"
+        The type of score to return. If "relevance", will return only the `_relevance_score. If "all", will return all scores from the vector and FTS search along with the relevance score.
+
+### Cohere Reranker
+This re-ranker uses the [Cohere](https://cohere.ai/) API to combine the results of semantic and full-text search. You can use this re-ranker by passing `CohereReranker()` to the `rerank()` method. Note that you'll need to set the `COHERE_API_KEY` environment variable to use this re-ranker.
+
+```python
+from lancedb.rerankers import CohereReranker
+
+reranker = CohereReranker()
+
+results = table.search("vampire weekend", query_type="hybrid").rerank(reranker=reranker).to_pandas()
+```
+
+### Arguments
+----------------
+* `model_name` : str, default `"rerank-english-v2.0"`
+        The name of the cross encoder model to use. Available cohere models are:
+        - rerank-english-v2.0
+        - rerank-multilingual-v2.0
+* `column` : str, default `"text"`
+        The name of the column to use as input to the cross encoder model.
+* `top_n` : str, default `None`
+        The number of results to return. If None, will return all results.
+
+!!! Note
+    Only returns `_relevance_score`. Does not support `return_score = "all"`.
+
+### Cross Encoder Reranker
+This reranker uses the [Sentence Transformers](https://www.sbert.net/) library to combine the results of semantic and full-text search. You can use it by passing `CrossEncoderReranker()` to the `rerank()` method.
+
+```python
+from lancedb.rerankers import CrossEncoderReranker
+
+reranker = CrossEncoderReranker()
+
+results = table.search("harmony hall", query_type="hybrid").rerank(reranker=reranker).to_pandas()
+```
+
+
+### Arguments
+----------------
+* `model` : str, default `"cross-encoder/ms-marco-TinyBERT-L-6"`
+        The name of the cross encoder model to use. Available cross encoder models can be found [here](https://www.sbert.net/docs/pretrained_cross-encoders.html)
+* `column` : str, default `"text"`
+        The name of the column to use as input to the cross encoder model.
+* `device` : str, default `None`
+        The device to use for the cross encoder model. If None, will use "cuda" if available, otherwise "cpu".
+
+!!! Note
+    Only returns `_relevance_score`. Does not support `return_score = "all"`.
+
+
+### ColBERT Reranker
+This reranker uses the ColBERT model to combine the results of semantic and full-text search. You can use it by passing `ColbertrReranker()` to the `rerank()` method. 
+
+ColBERT reranker model calculates relevance of given docs against the query and don't take existing fts and vector search scores into account, so it currently only supports `return_score="relevance"`. By default, it looks for `text` column to rerank the results. But you can specify the column name to use as input to the cross encoder model as described below.
+
+```python
+from lancedb.rerankers import ColbertReranker
+
+reranker = ColbertReranker()
+
+results = table.search("harmony hall", query_type="hybrid").rerank(reranker=reranker).to_pandas()
+```
+
+### Arguments
+----------------
+* `model_name` : `str`, default `"colbert-ir/colbertv2.0"`
+        The name of the cross encoder model to use.
+* `column` : `str`, default `"text"`
+        The name of the column to use as input to the cross encoder model.
+* `return_score` : `str`, default `"relevance"`
+        options are `"relevance"` or `"all"`. Only `"relevance"` is supported for now.
+
+!!! Note
+    Only returns `_relevance_score`. Does not support `return_score = "all"`.
+
+### OpenAI Reranker
+This reranker uses the OpenAI API to combine the results of semantic and full-text search. You can use it by passing `OpenaiReranker()` to the `rerank()` method.
+
+!!! Note
+    This prompts chat model to rerank results which is not a dedicated reranker model. This should be treated as experimental.
+
+!!! Tip
+    - You might run out of token limit so set the search `limits` based on your token limit.
+    - It is recommended to use gpt-4-turbo-preview, the default model, older models might lead to undesired behaviour
+
+```python
+from lancedb.rerankers import OpenaiReranker
+
+reranker = OpenaiReranker()
+
+results = table.search("harmony hall", query_type="hybrid").rerank(reranker=reranker).to_pandas()
+```
+
+### Arguments
+----------------
+* `model_name` : `str`, default `"gpt-4-turbo-preview"`
+    The name of the cross encoder model to use.
+* `column` : `str`, default `"text"`
+    The name of the column to use as input to the cross encoder model.
+* `return_score` : `str`, default `"relevance"`
+    options are "relevance" or "all". Only "relevance" is supported for now.
+* `api_key` : `str`, default `None`
+    The API key to use. If None, will use the OPENAI_API_KEY environment variable.
+
+
+## Building Custom Rerankers
+You can build your own custom reranker by subclassing the `Reranker` class and implementing the `rerank_hybrid()` method. Here's an example of a custom reranker that combines the results of semantic and full-text search using a linear combination of the scores.
+
+The `Reranker` base interface comes with a `merge_results()` method that can be used to combine the results of semantic and full-text search. This is a vanilla merging algorithm that simply concatenates the results and removes the duplicates without taking the scores into consideration. It only keeps the first copy of the row encountered. This works well in cases that don't require the scores of semantic and full-text search to combine the results. If you want to use the scores or want to support `return_score="all"`, you'll need to implement your own merging algorithm.
+
+```python
+
+from lancedb.rerankers import Reranker
+import pyarrow as pa
+
+class MyReranker(Reranker):
+    def __init__(self, param1, param2, ..., return_score="relevance"):
+        super().__init__(return_score)
+        self.param1 = param1
+        self.param2 = param2
+    
+    def rerank_hybrid(self, query: str, vector_results: pa.Table, fts_results: pa.Table):
+        # Use the built-in merging function
+        combined_result = self.merge_results(vector_results, fts_results)
+        
+        # Do something with the combined results
+        # ...
+
+        # Return the combined results
+        return combined_result
+
+```
+
+### Example of a Custom Reranker
+For the sake of simplicity let's build custom reranker that just enchances the Cohere Reranker by accepting a filter query, and accept other CohereReranker params as kwags.
+
+```python
+
+from typing import List, Union
+import pandas as pd
+from lancedb.rerankers import CohereReranker
+
+class MofidifiedCohereReranker(CohereReranker):
+    def __init__(self, filters: Union[str, List[str]], **kwargs):
+        super().__init__(**kwargs)
+        filters = filters if isinstance(filters, list) else [filters]
+        self.filters = filters
+
+    def rerank_hybrid(self, query: str, vector_results: pa.Table, fts_results: pa.Table)-> pa.Table:
+        combined_result = super().rerank_hybrid(query, vector_results, fts_results)
+        df = combined_result.to_pandas()
+        for filter in self.filters:
+            df = df.query("not text.str.contains(@filter)")
+
+        return pa.Table.from_pandas(df)
+
+```
+
+!!! tip
+    The `vector_results` and `fts_results` are pyarrow tables. You can convert them to pandas dataframes using `to_pandas()` method and perform any operations you want. After you are done, you can convert the dataframe back to pyarrow table using `pa.Table.from_pandas()` method and return it.

docs/src/notebooks/tables_guide.ipynb

@@ -13,7 +13,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 50,
+   "execution_count": 2,
    "id": "c1b4e34b-a49c-471d-a343-a5940bb5138a",
    "metadata": {},
    "outputs": [],
@@ -23,7 +23,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": 3,
    "id": "4e5a8d07-d9a1-48c1-913a-8e0629289579",
    "metadata": {},
    "outputs": [],
@@ -44,7 +44,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": 4,
    "id": "5df12f66-8d99-43ad-8d0b-22189ec0a6b9",
    "metadata": {},
    "outputs": [
@@ -62,7 +62,7 @@
        "long: [[-122.7,-74.1]]"
       ]
      },
-     "execution_count": 2,
+     "execution_count": 4,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -90,7 +90,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 3,
+   "execution_count": 5,
    "id": "f4d87ae9-0ccb-48eb-b31d-bb8f2370e47e",
    "metadata": {},
    "outputs": [
@@ -108,7 +108,7 @@
        "long: [[-122.7,-74.1]]"
       ]
      },
-     "execution_count": 3,
+     "execution_count": 5,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -135,10 +135,17 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 8,
+   "execution_count": 6,
    "id": "25f34bcf-fca0-4431-8601-eac95d1bd347",
    "metadata": {},
    "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "[2024-01-31T18:59:33Z WARN  lance::dataset] No existing dataset at /Users/qian/Work/LanceDB/lancedb/docs/src/notebooks/.lancedb/table3.lance, it will be created\n"
+     ]
+    },
     {
      "data": {
       "text/plain": [
@@ -148,7 +155,7 @@
        "long: float"
       ]
      },
-     "execution_count": 8,
+     "execution_count": 6,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -171,45 +178,51 @@
    "id": "4df51925-7ca2-4005-9c72-38b3d26240c6",
    "metadata": {},
    "source": [
-    "### From PyArrow Tables\n",
+    "### From an Arrow Table\n",
     "\n",
-    "You can also create LanceDB tables directly from pyarrow tables"
+    "You can also create LanceDB tables directly from pyarrow tables. LanceDB supports float16 type."
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 12,
+   "execution_count": 7,
    "id": "90a880f6-be43-4c9d-ba65-0b05197c0f6f",
    "metadata": {},
    "outputs": [
     {
      "data": {
       "text/plain": [
-       "vector: fixed_size_list<item: float>[2]\n",
-       "  child 0, item: float\n",
-       "item: string\n",
-       "price: double"
+       "vector: fixed_size_list<item: halffloat>[16]\n",
+       "  child 0, item: halffloat\n",
+       "text: string"
       ]
      },
-     "execution_count": 12,
+     "execution_count": 7,
      "metadata": {},
      "output_type": "execute_result"
     }
    ],
    "source": [
-    "table = pa.Table.from_arrays(\n",
-    "        [\n",
-    "            pa.array([[3.1, 4.1], [5.9, 26.5]],\n",
-    "                    pa.list_(pa.float32(), 2)),\n",
-    "            pa.array([\"foo\", \"bar\"]),\n",
-    "            pa.array([10.0, 20.0]),\n",
-    "        ],\n",
-    "        [\"vector\", \"item\", \"price\"],\n",
-    "    )\n",
+    "import numpy as np\n",
     "\n",
-    "db = lancedb.connect(\"db\")\n",
+    "dim = 16\n",
+    "total = 2\n",
+    "schema = pa.schema(\n",
+    "    [\n",
+    "        pa.field(\"vector\", pa.list_(pa.float16(), dim)),\n",
+    "        pa.field(\"text\", pa.string())\n",
+    "    ]\n",
+    ")\n",
+    "data = pa.Table.from_arrays(\n",
+    "    [\n",
+    "        pa.array([np.random.randn(dim).astype(np.float16) for _ in range(total)],\n",
+    "                pa.list_(pa.float16(), dim)),\n",
+    "        pa.array([\"foo\", \"bar\"])\n",
+    "    ],\n",
+    "    [\"vector\", \"text\"],\n",
+    ")\n",
     "\n",
-    "tbl = db.create_table(\"test1\", table, mode=\"overwrite\")\n",
+    "tbl = db.create_table(\"f16_tbl\", data, schema=schema)\n",
     "tbl.schema"
    ]
   },
@@ -225,7 +238,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 13,
+   "execution_count": 8,
    "id": "d81121d7-e4b7-447c-a48c-974b6ebb464a",
    "metadata": {},
    "outputs": [
@@ -240,7 +253,7 @@
        "imdb_id: int64 not null"
       ]
      },
-     "execution_count": 13,
+     "execution_count": 8,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -282,7 +295,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 14,
+   "execution_count": 9,
    "id": "bc247142-4e3c-41a2-b94c-8e00d2c2a508",
    "metadata": {},
    "outputs": [
@@ -292,7 +305,7 @@
        "LanceTable(table4)"
       ]
      },
-     "execution_count": 14,
+     "execution_count": 9,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -333,7 +346,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 16,
+   "execution_count": 10,
    "id": "25ad3523-e0c9-4c28-b3df-38189c4e0e5f",
    "metadata": {},
    "outputs": [
@@ -346,7 +359,7 @@
        "price: double not null"
       ]
      },
-     "execution_count": 16,
+     "execution_count": 10,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -385,7 +398,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 17,
+   "execution_count": 11,
    "id": "2814173a-eacc-4dd8-a64d-6312b44582cc",
    "metadata": {},
    "outputs": [],
@@ -411,7 +424,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 18,
+   "execution_count": 12,
    "id": "df9e13c0-41f6-437f-9dfa-2fd71d3d9c45",
    "metadata": {},
    "outputs": [
@@ -421,7 +434,7 @@
        "['table6', 'table4', 'table5', 'movielens_small']"
       ]
      },
-     "execution_count": 18,
+     "execution_count": 12,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -432,7 +445,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 20,
+   "execution_count": 13,
    "id": "9343f5ad-6024-42ee-ac2f-6c1471df8679",
    "metadata": {},
    "outputs": [
@@ -541,7 +554,7 @@
        "9  [5.9, 26.5]  bar   20.0"
       ]
      },
-     "execution_count": 20,
+     "execution_count": 13,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -564,7 +577,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 21,
+   "execution_count": 14,
    "id": "8a56250f-73a1-4c26-a6ad-5c7a0ce3a9ab",
    "metadata": {},
    "outputs": [],
@@ -590,7 +603,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 22,
+   "execution_count": 15,
    "id": "030c7057-b98e-4e2f-be14-b8c1f927f83c",
    "metadata": {},
    "outputs": [],
@@ -621,7 +634,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 24,
+   "execution_count": 16,
    "id": "e7a17de2-08d2-41b7-bd05-f63d1045ab1f",
    "metadata": {},
    "outputs": [
@@ -629,16 +642,16 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "32\n"
+      "22\n"
      ]
     },
     {
      "data": {
       "text/plain": [
-       "17"
+       "12"
       ]
      },
-     "execution_count": 24,
+     "execution_count": 16,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -661,7 +674,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 30,
+   "execution_count": 17,
    "id": "fe3310bd-08f4-4a22-a63b-b3127d22f9f7",
    "metadata": {},
    "outputs": [
@@ -681,25 +694,20 @@
       "8    [3.1, 4.1]   foo   10.0\n",
       "9    [3.1, 4.1]   foo   10.0\n",
       "10   [3.1, 4.1]   foo   10.0\n",
-      "11   [3.1, 4.1]   foo   10.0\n",
-      "12   [3.1, 4.1]   foo   10.0\n",
-      "13   [3.1, 4.1]   foo   10.0\n",
-      "14   [3.1, 4.1]   foo   10.0\n",
-      "15   [3.1, 4.1]   foo   10.0\n",
-      "16   [3.1, 4.1]   foo   10.0\n"
+      "11   [3.1, 4.1]   foo   10.0\n"
      ]
     },
     {
      "ename": "OSError",
-     "evalue": "LanceError(IO): Error during planning: column foo does not exist",
+     "evalue": "LanceError(IO): Error during planning: column foo does not exist, /Users/runner/work/lance/lance/rust/lance-core/src/error.rs:212:23",
      "output_type": "error",
      "traceback": [
       "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
       "\u001b[0;31mOSError\u001b[0m                                   Traceback (most recent call last)",
-      "Cell \u001b[0;32mIn[30], line 4\u001b[0m\n\u001b[1;32m      2\u001b[0m to_remove \u001b[38;5;241m=\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124m, \u001b[39m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;241m.\u001b[39mjoin(\u001b[38;5;28mstr\u001b[39m(v) \u001b[38;5;28;01mfor\u001b[39;00m v \u001b[38;5;129;01min\u001b[39;00m to_remove)\n\u001b[1;32m      3\u001b[0m \u001b[38;5;28mprint\u001b[39m(tbl\u001b[38;5;241m.\u001b[39mto_pandas())\n\u001b[0;32m----> 4\u001b[0m \u001b[43mtbl\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mdelete\u001b[49m\u001b[43m(\u001b[49m\u001b[38;5;124;43mf\u001b[39;49m\u001b[38;5;124;43m\"\u001b[39;49m\u001b[38;5;124;43mitem IN (\u001b[39;49m\u001b[38;5;132;43;01m{\u001b[39;49;00m\u001b[43mto_remove\u001b[49m\u001b[38;5;132;43;01m}\u001b[39;49;00m\u001b[38;5;124;43m)\u001b[39;49m\u001b[38;5;124;43m\"\u001b[39;49m\u001b[43m)\u001b[49m\n\u001b[1;32m      5\u001b[0m tbl\u001b[38;5;241m.\u001b[39mto_pandas()\n",
-      "File \u001b[0;32m~/Documents/lancedb/lancedb/python/lancedb/table.py:610\u001b[0m, in \u001b[0;36mLanceTable.delete\u001b[0;34m(self, where)\u001b[0m\n\u001b[1;32m    609\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21mdelete\u001b[39m(\u001b[38;5;28mself\u001b[39m, where: \u001b[38;5;28mstr\u001b[39m):\n\u001b[0;32m--> 610\u001b[0m     \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43m_dataset\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mdelete\u001b[49m\u001b[43m(\u001b[49m\u001b[43mwhere\u001b[49m\u001b[43m)\u001b[49m\n",
-      "File \u001b[0;32m~/Documents/lancedb/lancedb/env/lib/python3.11/site-packages/lance/dataset.py:489\u001b[0m, in \u001b[0;36mLanceDataset.delete\u001b[0;34m(self, predicate)\u001b[0m\n\u001b[1;32m    487\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;28misinstance\u001b[39m(predicate, pa\u001b[38;5;241m.\u001b[39mcompute\u001b[38;5;241m.\u001b[39mExpression):\n\u001b[1;32m    488\u001b[0m     predicate \u001b[38;5;241m=\u001b[39m \u001b[38;5;28mstr\u001b[39m(predicate)\n\u001b[0;32m--> 489\u001b[0m \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43m_ds\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mdelete\u001b[49m\u001b[43m(\u001b[49m\u001b[43mpredicate\u001b[49m\u001b[43m)\u001b[49m\n",
-      "\u001b[0;31mOSError\u001b[0m: LanceError(IO): Error during planning: column foo does not exist"
+      "Cell \u001b[0;32mIn[17], line 4\u001b[0m\n\u001b[1;32m      2\u001b[0m to_remove \u001b[38;5;241m=\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124m, \u001b[39m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;241m.\u001b[39mjoin(\u001b[38;5;28mstr\u001b[39m(v) \u001b[38;5;28;01mfor\u001b[39;00m v \u001b[38;5;129;01min\u001b[39;00m to_remove)\n\u001b[1;32m      3\u001b[0m \u001b[38;5;28mprint\u001b[39m(tbl\u001b[38;5;241m.\u001b[39mto_pandas())\n\u001b[0;32m----> 4\u001b[0m \u001b[43mtbl\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mdelete\u001b[49m\u001b[43m(\u001b[49m\u001b[38;5;124;43mf\u001b[39;49m\u001b[38;5;124;43m\"\u001b[39;49m\u001b[38;5;124;43mitem IN (\u001b[39;49m\u001b[38;5;132;43;01m{\u001b[39;49;00m\u001b[43mto_remove\u001b[49m\u001b[38;5;132;43;01m}\u001b[39;49;00m\u001b[38;5;124;43m)\u001b[39;49m\u001b[38;5;124;43m\"\u001b[39;49m\u001b[43m)\u001b[49m\n",
+      "File \u001b[0;32m~/Work/LanceDB/lancedb/docs/doc-venv/lib/python3.11/site-packages/lancedb/table.py:872\u001b[0m, in \u001b[0;36mLanceTable.delete\u001b[0;34m(self, where)\u001b[0m\n\u001b[1;32m    871\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21mdelete\u001b[39m(\u001b[38;5;28mself\u001b[39m, where: \u001b[38;5;28mstr\u001b[39m):\n\u001b[0;32m--> 872\u001b[0m     \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43m_dataset\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mdelete\u001b[49m\u001b[43m(\u001b[49m\u001b[43mwhere\u001b[49m\u001b[43m)\u001b[49m\n",
+      "File \u001b[0;32m~/Work/LanceDB/lancedb/docs/doc-venv/lib/python3.11/site-packages/lance/dataset.py:596\u001b[0m, in \u001b[0;36mLanceDataset.delete\u001b[0;34m(self, predicate)\u001b[0m\n\u001b[1;32m    594\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;28misinstance\u001b[39m(predicate, pa\u001b[38;5;241m.\u001b[39mcompute\u001b[38;5;241m.\u001b[39mExpression):\n\u001b[1;32m    595\u001b[0m     predicate \u001b[38;5;241m=\u001b[39m \u001b[38;5;28mstr\u001b[39m(predicate)\n\u001b[0;32m--> 596\u001b[0m \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43m_ds\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mdelete\u001b[49m\u001b[43m(\u001b[49m\u001b[43mpredicate\u001b[49m\u001b[43m)\u001b[49m\n",
+      "\u001b[0;31mOSError\u001b[0m: LanceError(IO): Error during planning: column foo does not exist, /Users/runner/work/lance/lance/rust/lance-core/src/error.rs:212:23"
      ]
     }
    ],
@@ -712,7 +720,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 43,
+   "execution_count": null,
    "id": "87d5bc21-847f-4c81-b56e-f6dbe5d05aac",
    "metadata": {},
    "outputs": [],
@@ -729,7 +737,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 44,
+   "execution_count": null,
    "id": "9cba4519-eb3a-4941-ab7e-873d762e750f",
    "metadata": {},
    "outputs": [],
@@ -742,7 +750,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 46,
+   "execution_count": null,
    "id": "5bdc9801-d5ed-4871-92d0-88b27108e788",
    "metadata": {},
    "outputs": [
@@ -817,7 +825,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.11.4"
+   "version": "3.11.7"
   }
  },
  "nbformat": 4,

docs/src/python/python.md

@@ -58,6 +58,8 @@ pip install lancedb
 
 ::: lancedb.schema.vector
 
+::: lancedb.merge.LanceMergeInsertBuilder
+
 ## Integrations
 
 ### Pydantic

docs/src/search.md

@@ -2,27 +2,26 @@
 
 A vector search finds the approximate or exact nearest neighbors to a given query vector.
 
-* In a recommendation system or search engine, you can find similar records to
-the one you searched.
-* In LLM and other AI applications,
-each data point can be represented by [embeddings generated from existing models](embeddings/index.md),
-following which the search returns the most relevant features.
+- In a recommendation system or search engine, you can find similar records to
+  the one you searched.
+- In LLM and other AI applications,
+  each data point can be represented by [embeddings generated from existing models](embeddings/index.md),
+  following which the search returns the most relevant features.
 
 ## Distance metrics
 
 Distance metrics are a measure of the similarity between a pair of vectors.
 Currently, LanceDB supports the following metrics:
 
-| Metric      | Description                          |
-| ----------- | ------------------------------------ |
-| `l2`        | [Euclidean / L2 distance](https://en.wikipedia.org/wiki/Euclidean_distance) |
-| `cosine`    | [Cosine Similarity](https://en.wikipedia.org/wiki/Cosine_similarity)|
-| `dot`       | [Dot Production](https://en.wikipedia.org/wiki/Dot_product) |
-
+| Metric   | Description                                                                 |
+| -------- | --------------------------------------------------------------------------- |
+| `l2`     | [Euclidean / L2 distance](https://en.wikipedia.org/wiki/Euclidean_distance) |
+| `cosine` | [Cosine Similarity](https://en.wikipedia.org/wiki/Cosine_similarity)        |
+| `dot`    | [Dot Production](https://en.wikipedia.org/wiki/Dot_product)                 |
 
 ## Exhaustive search (kNN)
 
-If you do not create a vector index, LanceDB exhaustively scans the *entire* vector space
+If you do not create a vector index, LanceDB exhaustively scans the _entire_ vector space
 and compute the distance to every vector in order to find the exact nearest neighbors. This is effectively a kNN search.
 
 <!-- Setup Code
@@ -38,22 +37,9 @@ data = [{"vector": row, "item": f"item {i}"}
 db.create_table("my_vectors", data=data)
 ```
 -->
-<!-- Setup Code
-```javascript
-const vectordb_setup = require('vectordb')
-const db_setup = await vectordb_setup.connect('data/sample-lancedb')
-
-let data = []
-for (let i = 0; i < 10_000; i++) {
-     data.push({vector: Array(1536).fill(i), id: `${i}`, content: "", longId: `${i}`},)
-}
-await db_setup.createTable('my_vectors', data)
-```
--->
 
 === "Python"
 
-
     ```python
     import lancedb
     import numpy as np
@@ -70,17 +56,12 @@ await db_setup.createTable('my_vectors', data)
 === "JavaScript"
 
     ```javascript
-    const vectordb = require('vectordb')
-    const db = await vectordb.connect('data/sample-lancedb')
+    --8<-- "docs/src/search_legacy.ts:import"
 
-    const tbl = await db.openTable("my_vectors")
-
-    const results_1 = await tbl.search(Array(1536).fill(1.2))
-        .limit(10)
-        .execute()
+    --8<-- "docs/src/search_legacy.ts:search1"
     ```
 
-By default, `l2` will be used as metric type. You can specify the metric type as 
+By default, `l2` will be used as metric type. You can specify the metric type as
 `cosine` or `dot` if required.
 
 === "Python"
@@ -92,20 +73,16 @@ By default, `l2` will be used as metric type. You can specify the metric type as
         .to_list()
     ```
 
-
 === "JavaScript"
 
     ```javascript
-    const results_2 = await tbl.search(Array(1536).fill(1.2))
-        .metricType("cosine")
-        .limit(10)
-        .execute()
+    --8<-- "docs/src/search_legacy.ts:search2"
     ```
 
 ## Approximate nearest neighbor (ANN) search
 
 To perform scalable vector retrieval with acceptable latencies, it's common to build a vector index.
-While the exhaustive search is guaranteed to always return 100% recall, the approximate nature of 
+While the exhaustive search is guaranteed to always return 100% recall, the approximate nature of
 an ANN search means that using an index often involves a trade-off between recall and latency.
 
 See the [IVF_PQ index](./concepts/index_ivfpq.md.md) for a deeper description of how `IVF_PQ`
@@ -117,7 +94,9 @@ LanceDB returns vector search results via different formats commonly used in pyt
 Let's create a LanceDB table with a nested schema:
 
 === "Python"
+
     ```python
+
     from datetime import datetime
     import lancedb
     from lancedb.pydantic import LanceModel, Vector
@@ -153,7 +132,7 @@ Let's create a LanceDB table with a nested schema:
     ### As a PyArrow table
 
     Using `to_arrow()` we can get the results back as a pyarrow Table.
-    This result table has the same columns as the LanceDB table, with 
+    This result table has the same columns as the LanceDB table, with
     the addition of an `_distance` column for vector search or a `score`
     column for full text search.
 
@@ -169,11 +148,11 @@ Let's create a LanceDB table with a nested schema:
     tbl.search(np.random.randn(1536)).to_pandas()
     ```
 
-    While other formats like Arrow/Pydantic/Python dicts have a natural 
-    way to handle nested schemas, pandas can only store nested data as a 
+    While other formats like Arrow/Pydantic/Python dicts have a natural
+    way to handle nested schemas, pandas can only store nested data as a
     python dict column, which makes it difficult to support nested references.
-    So for convenience, you can also tell LanceDB to flatten a nested schema 
-    when creating the pandas dataframe. 
+    So for convenience, you can also tell LanceDB to flatten a nested schema
+    when creating the pandas dataframe.
 
     ```python
     tbl.search(np.random.randn(1536)).to_pandas(flatten=True)

docs/src/search_legacy.ts

@@ -0,0 +1,41 @@
+// --8<-- [start:import]
+import * as lancedb from "vectordb";
+// --8<-- [end:import]
+import * as fs from "fs";
+
+async function setup() {
+  fs.rmSync("data/sample-lancedb", { recursive: true, force: true });
+  const db = await lancedb.connect("data/sample-lancedb");
+
+  let data = [];
+  for (let i = 0; i < 10_000; i++) {
+    data.push({
+      vector: Array(1536).fill(i),
+      id: `${i}`,
+      content: "",
+      longId: `${i}`,
+    });
+  }
+  await db.createTable("my_vectors", data);
+}
+
+async () => {
+  await setup();
+
+  // --8<-- [start:search1]
+  const db = await lancedb.connect("data/sample-lancedb");
+  const tbl = await db.openTable("my_vectors");
+
+  const results_1 = await tbl.search(Array(1536).fill(1.2)).limit(10).execute();
+  // --8<-- [end:search1]
+
+  // --8<-- [start:search2]
+  const results_2 = await tbl
+    .search(Array(1536).fill(1.2))
+    .metricType(lancedb.MetricType.Cosine)
+    .limit(10)
+    .execute();
+  // --8<-- [end:search2]
+
+  console.log("search: done");
+};

docs/src/sql.md

@@ -8,7 +8,7 @@ option that performs the filter prior to vector search. This can be useful to na
 the search space on a very large dataset to reduce query latency.
 
 <!-- Setup Code
-```python 
+```python
 import lancedb
 import numpy as np
 uri = "data/sample-lancedb"
@@ -21,7 +21,7 @@ tbl = db.create_table("my_vectors", data=data)
 ```
 -->
 <!-- Setup Code
-```javascript 
+```javascript
 const vectordb = require('vectordb')
 const db = await vectordb.connect('data/sample-lancedb')
 
@@ -34,6 +34,7 @@ const tbl = await db.createTable('myVectors', data)
 -->
 
 === "Python"
+
     ```py
     result = (
         tbl.search([0.5, 0.2])
@@ -44,12 +45,9 @@ const tbl = await db.createTable('myVectors', data)
     ```
 
 === "JavaScript"
+
     ```javascript
-    let result = await tbl.search(Array(1536).fill(0.5))
-        .limit(1)
-        .filter("id = 10")
-        .prefilter(true)
-        .execute()
+    --8<-- "docs/src/sql_legacy.ts:search"
     ```
 
 ## SQL filters
@@ -60,14 +58,14 @@ It can be used during vector search, update, and deletion operations.
 
 Currently, Lance supports a growing list of SQL expressions.
 
-* ``>``, ``>=``, ``<``, ``<=``, ``=``
-* ``AND``, ``OR``, ``NOT``
-* ``IS NULL``, ``IS NOT NULL``
-* ``IS TRUE``, ``IS NOT TRUE``, ``IS FALSE``, ``IS NOT FALSE``
-* ``IN``
-* ``LIKE``, ``NOT LIKE``
-* ``CAST``
-* ``regexp_match(column, pattern)``
+- `>`, `>=`, `<`, `<=`, `=`
+- `AND`, `OR`, `NOT`
+- `IS NULL`, `IS NOT NULL`
+- `IS TRUE`, `IS NOT TRUE`, `IS FALSE`, `IS NOT FALSE`
+- `IN`
+- `LIKE`, `NOT LIKE`
+- `CAST`
+- `regexp_match(column, pattern)`
 
 For example, the following filter string is acceptable:
 
@@ -82,29 +80,27 @@ For example, the following filter string is acceptable:
 === "Javascript"
 
     ```javascript
-    await tbl.search(Array(1536).fill(0))
-       .where("(item IN ('item 0', 'item 2')) AND (id > 10)")
-       .execute()
+    --8<-- "docs/src/sql_legacy.ts:vec_search"
     ```
 
-
 If your column name contains special characters or is a [SQL Keyword](https://docs.rs/sqlparser/latest/sqlparser/keywords/index.html),
 you can use backtick (`` ` ``) to escape it. For nested fields, each segment of the
 path must be wrapped in backticks.
 
 === "SQL"
+
     ```sql
     `CUBE` = 10 AND `column name with space` IS NOT NULL
       AND `nested with space`.`inner with space` < 2
     ```
 
-!!! warning
-    Field names containing periods (``.``) are not supported.
+!!!warning "Field names containing periods (`.`) are not supported."
 
 Literals for dates, timestamps, and decimals can be written by writing the string
 value after the type name. For example
 
 === "SQL"
+
     ```sql
     date_col = date '2021-01-01'
     and timestamp_col = timestamp '2021-01-01 00:00:00'
@@ -114,49 +110,47 @@ value after the type name. For example
 For timestamp columns, the precision can be specified as a number in the type
 parameter. Microsecond precision (6) is the default.
 
-| SQL              | Time unit    |
-|------------------|--------------|
-| ``timestamp(0)`` | Seconds      |
-| ``timestamp(3)`` | Milliseconds |
-| ``timestamp(6)`` | Microseconds |
-| ``timestamp(9)`` | Nanoseconds  |
+| SQL            | Time unit    |
+| -------------- | ------------ |
+| `timestamp(0)` | Seconds      |
+| `timestamp(3)` | Milliseconds |
+| `timestamp(6)` | Microseconds |
+| `timestamp(9)` | Nanoseconds  |
 
 LanceDB internally stores data in [Apache Arrow](https://arrow.apache.org/) format.
 The mapping from SQL types to Arrow types is:
 
-| SQL type | Arrow type |
-|----------|------------|
-| ``boolean`` | ``Boolean`` |
-| ``tinyint`` / ``tinyint unsigned`` | ``Int8`` / ``UInt8`` |
-| ``smallint`` / ``smallint unsigned`` | ``Int16`` / ``UInt16`` |
-| ``int`` or ``integer`` / ``int unsigned`` or ``integer unsigned`` | ``Int32`` / ``UInt32`` |
-| ``bigint`` / ``bigint unsigned`` | ``Int64`` / ``UInt64`` |
-| ``float`` | ``Float32`` |
-| ``double`` | ``Float64`` |
-| ``decimal(precision, scale)`` | ``Decimal128`` |
-| ``date`` | ``Date32`` |
-| ``timestamp`` | ``Timestamp`` [^1] |
-| ``string`` | ``Utf8`` |
-| ``binary`` | ``Binary`` |
+| SQL type                                                  | Arrow type         |
+| --------------------------------------------------------- | ------------------ |
+| `boolean`                                                 | `Boolean`          |
+| `tinyint` / `tinyint unsigned`                            | `Int8` / `UInt8`   |
+| `smallint` / `smallint unsigned`                          | `Int16` / `UInt16` |
+| `int` or `integer` / `int unsigned` or `integer unsigned` | `Int32` / `UInt32` |
+| `bigint` / `bigint unsigned`                              | `Int64` / `UInt64` |
+| `float`                                                   | `Float32`          |
+| `double`                                                  | `Float64`          |
+| `decimal(precision, scale)`                               | `Decimal128`       |
+| `date`                                                    | `Date32`           |
+| `timestamp`                                               | `Timestamp` [^1]   |
+| `string`                                                  | `Utf8`             |
+| `binary`                                                  | `Binary`           |
 
 [^1]: See precision mapping in previous table.
 
-
 ## Filtering without Vector Search
 
 You can also filter your data without search.
 
 === "Python"
-      ```python
-      tbl.search().where("id = 10").limit(10).to_arrow()
-      ```
+
+    ```python
+    tbl.search().where("id = 10").limit(10).to_arrow()
+    ```
 
 === "JavaScript"
-      ```javascript
-      await tbl.where('id = 10').limit(10).execute()
-      ```
 
-!!! warning
-    If your table is large, this could potentially return a very large
-    amount of data. Please be sure to use a `limit` clause unless
-    you're sure you want to return the whole result set.
+    ```javascript
+    --8<---- "docs/src/sql_legacy.ts:sql_search"
+    ```
+
+!!!warning "If your table is large, this could potentially return a very large amount of data. Please be sure to use a `limit` clause unless you're sure you want to return the whole result set."

docs/src/sql_legacy.ts

@@ -0,0 +1,38 @@
+import * as vectordb from "vectordb";
+
+(async () => {
+  const db = await vectordb.connect("data/sample-lancedb");
+
+  let data = [];
+  for (let i = 0; i < 10_000; i++) {
+    data.push({
+      vector: Array(1536).fill(i),
+      id: i,
+      item: `item ${i}`,
+      strId: `${i}`,
+    });
+  }
+  const tbl = await db.createTable("myVectors", data);
+
+  // --8<-- [start:search]
+  let result = await tbl
+    .search(Array(1536).fill(0.5))
+    .limit(1)
+    .filter("id = 10")
+    .prefilter(true)
+    .execute();
+  // --8<-- [end:search]
+
+  // --8<-- [start:vec_search]
+  await tbl
+    .search(Array(1536).fill(0))
+    .where("(item IN ('item 0', 'item 2')) AND (id > 10)")
+    .execute();
+  // --8<-- [end:vec_search]
+
+  // --8<-- [start:sql_search]
+  await tbl.filter("id = 10").limit(10).execute();
+  // --8<-- [end:sql_search]
+
+  console.log("SQL search: done");
+})();

docs/test/md_testing.js

@@ -1,54 +0,0 @@
-const glob = require("glob");
-const fs = require("fs");
-const path = require("path");
-
-const globString = "../src/**/*.md";
-
-const excludedGlobs = [
-  "../src/fts.md",
-  "../src/embedding.md",
-  "../src/examples/*.md",
-  "../src/guides/tables.md",
-  "../src/embeddings/*.md",
-];
-
-const nodePrefix = "javascript";
-const nodeFile = ".js";
-const nodeFolder = "node";
-const asyncPrefix = "(async () => {\n";
-const asyncSuffix = "})();";
-
-function* yieldLines(lines, prefix, suffix) {
-  let inCodeBlock = false;
-  for (const line of lines) {
-    if (line.trim().startsWith(prefix + nodePrefix)) {
-      inCodeBlock = true;
-    } else if (inCodeBlock && line.trim().startsWith(suffix)) {
-      inCodeBlock = false;
-      yield "\n";
-    } else if (inCodeBlock) {
-      yield line;
-    }
-  }
-}
-
-const files = glob.sync(globString, { recursive: true });
-const excludedFiles = glob.sync(excludedGlobs, { recursive: true });
-
-for (const file of files.filter((file) => !excludedFiles.includes(file))) {
-  const lines = [];
-  const data = fs.readFileSync(file, "utf-8");
-  const fileLines = data.split("\n");
-
-  for (const line of yieldLines(fileLines, "```", "```")) {
-    lines.push(line);
-  }
-
-  if (lines.length > 0) {
-    const fileName = path.basename(file, ".md");
-    const outPath = path.join(nodeFolder, fileName, `${fileName}${nodeFile}`);
-    console.log(outPath)
-    fs.mkdirSync(path.dirname(outPath), { recursive: true });
-    fs.writeFileSync(outPath, asyncPrefix + "\n" + lines.join("\n") + asyncSuffix);
-  }
-}

docs/test/md_testing.py

@@ -14,6 +14,7 @@ excluded_globs = [
     "../src/concepts/*.md",
     "../src/ann_indexes.md",
     "../src/basic.md",
+    "../src/hybrid_search/hybrid_search.md",
 ]
 
 python_prefix = "py"
@@ -48,6 +49,7 @@ def yield_lines(lines: Iterator[str], prefix: str, suffix: str):
             if not skip_test:
                 yield line[strip_length:]
 
+
 for file in filter(lambda file: file not in excluded_files, files):
     with open(file, "r") as f:
         lines = list(yield_lines(iter(f), "```", "```"))

docs/test/package.json

@@ -1,13 +0,0 @@
-{
-  "name": "lancedb-docs-test",
-  "version": "1.0.0",
-  "description": "",
-  "author": "",
-  "license": "ISC",
-  "dependencies": {
-    "fs": "^0.0.1-security",
-    "glob": "^10.2.7",
-    "path": "^0.12.7",
-    "vectordb": "https://gitpkg.now.sh/lancedb/lancedb/node?main"
-  }
-}

docs/tsconfig.json

@@ -0,0 +1,17 @@
+{
+  "include": [
+    "src/*.ts",
+  ],
+  "compilerOptions": {
+    "target": "es2022",
+    "module": "nodenext",
+    "declaration": true,
+    "outDir": "./dist",
+    "strict": true,
+    "allowJs": true,
+    "resolveJsonModule": true,
+  },
+  "exclude": [
+    "./dist/*",
+  ]
+}

node/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "vectordb",
-  "version": "0.4.4",
+  "version": "0.4.10",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "vectordb",
-      "version": "0.4.4",
+      "version": "0.4.10",
       "cpu": [
         "x64",
         "arm64"
@@ -53,11 +53,11 @@
         "uuid": "^9.0.0"
       },
       "optionalDependencies": {
-        "@lancedb/vectordb-darwin-arm64": "0.4.4",
-        "@lancedb/vectordb-darwin-x64": "0.4.4",
-        "@lancedb/vectordb-linux-arm64-gnu": "0.4.4",
-        "@lancedb/vectordb-linux-x64-gnu": "0.4.4",
-        "@lancedb/vectordb-win32-x64-msvc": "0.4.4"
+        "@lancedb/vectordb-darwin-arm64": "0.4.10",
+        "@lancedb/vectordb-darwin-x64": "0.4.10",
+        "@lancedb/vectordb-linux-arm64-gnu": "0.4.10",
+        "@lancedb/vectordb-linux-x64-gnu": "0.4.10",
+        "@lancedb/vectordb-win32-x64-msvc": "0.4.10"
       }
     },
     "node_modules/@75lb/deep-merge": {

node/package.json

@@ -1,12 +1,12 @@
 {
   "name": "vectordb",
-  "version": "0.4.4",
+  "version": "0.4.10",
   "description": " Serverless, low-latency vector database for AI applications",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
   "scripts": {
     "tsc": "tsc -b",
-    "build": "cargo-cp-artifact --artifact cdylib vectordb-node index.node -- cargo build --message-format=json && tsc -b",
+    "build": "npm run tsc && cargo-cp-artifact --artifact cdylib vectordb-node index.node -- cargo build --message-format=json",
     "build-release": "npm run build -- --release",
     "test": "npm run tsc && mocha -recursive dist/test",
     "integration-test": "npm run tsc && mocha -recursive dist/integration_test",
@@ -17,7 +17,11 @@
   },
   "repository": {
     "type": "git",
-    "url": "https://github.com/lancedb/lancedb/node"
+    "url": "https://github.com/lancedb/lancedb.git"
+  },
+  "homepage": "https://lancedb.github.io/lancedb/",
+  "bugs": {
+    "url": "https://github.com/lancedb/lancedb/issues"
   },
   "keywords": [
     "data-format",
@@ -81,10 +85,10 @@
     }
   },
   "optionalDependencies": {
-    "@lancedb/vectordb-darwin-arm64": "0.4.4",
-    "@lancedb/vectordb-darwin-x64": "0.4.4",
-    "@lancedb/vectordb-linux-arm64-gnu": "0.4.4",
-    "@lancedb/vectordb-linux-x64-gnu": "0.4.4",
-    "@lancedb/vectordb-win32-x64-msvc": "0.4.4"
+    "@lancedb/vectordb-darwin-arm64": "0.4.10",
+    "@lancedb/vectordb-darwin-x64": "0.4.10",
+    "@lancedb/vectordb-linux-arm64-gnu": "0.4.10",
+    "@lancedb/vectordb-linux-x64-gnu": "0.4.10",
+    "@lancedb/vectordb-win32-x64-msvc": "0.4.10"
   }
 }

node/src/arrow.ts

@@ -14,8 +14,6 @@
 
 import {
   Field,
-  type FixedSizeListBuilder,
-  Float32,
   makeBuilder,
   RecordBatchFileWriter,
   Utf8,
@@ -26,14 +24,19 @@ import {
   Table as ArrowTable,
   RecordBatchStreamWriter,
   List,
-  Float64,
   RecordBatch,
   makeData,
   Struct,
-  type Float
+  type Float,
+  DataType,
+  Binary,
+  Float32
 } from 'apache-arrow'
 import { type EmbeddingFunction } from './index'
 
+/*
+ * Options to control how a column should be converted to a vector array
+ */
 export class VectorColumnOptions {
   /** Vector column type. */
   type: Float = new Float32()
@@ -45,14 +48,50 @@ export class VectorColumnOptions {
 
 /** Options to control the makeArrowTable call. */
 export class MakeArrowTableOptions {
-  /** Provided schema. */
+  /*
+   * Schema of the data.
+   *
+   * If this is not provided then the data type will be inferred from the
+   * JS type.  Integer numbers will become int64, floating point numbers
+   * will become float64 and arrays will become variable sized lists with
+   * the data type inferred from the first element in the array.
+   *
+   * The schema must be specified if there are no records (e.g. to make
+   * an empty table)
+   */
   schema?: Schema
 
-  /** Vector columns */
+  /*
+   * Mapping from vector column name to expected type
+   *
+   * Lance expects vector columns to be fixed size list arrays (i.e. tensors)
+   * However, `makeArrowTable` will not infer this by default (it creates
+   * variable size list arrays).  This field can be used to indicate that a column
+   * should be treated as a vector column and converted to a fixed size list.
+   *
+   * The keys should be the names of the vector columns.  The value specifies the
+   * expected data type of the vector columns.
+   *
+   * If `schema` is provided then this field is ignored.
+   *
+   * By default, the column named "vector" will be assumed to be a float32
+   * vector column.
+   */
   vectorColumns: Record<string, VectorColumnOptions> = {
     vector: new VectorColumnOptions()
   }
 
+  /**
+   * If true then string columns will be encoded with dictionary encoding
+   *
+   * Set this to true if your string columns tend to repeat the same values
+   * often.  For more precise control use the `schema` property to specify the
+   * data type for individual columns.
+   *
+   * If `schema` is provided then this property is ignored.
+   */
+  dictionaryEncodeStrings: boolean = false
+
   constructor (values?: Partial<MakeArrowTableOptions>) {
     Object.assign(this, values)
   }
@@ -62,8 +101,29 @@ export class MakeArrowTableOptions {
  * An enhanced version of the {@link makeTable} function from Apache Arrow
  * that supports nested fields and embeddings columns.
  *
+ * This function converts an array of Record<String, any> (row-major JS objects)
+ * to an Arrow Table (a columnar structure)
+ *
  * Note that it currently does not support nulls.
  *
+ * If a schema is provided then it will be used to determine the resulting array
+ * types.  Fields will also be reordered to fit the order defined by the schema.
+ *
+ * If a schema is not provided then the types will be inferred and the field order
+ * will be controlled by the order of properties in the first record.
+ *
+ * If the input is empty then a schema must be provided to create an empty table.
+ *
+ * When a schema is not specified then data types will be inferred.  The inference
+ * rules are as follows:
+ *
+ *  - boolean => Bool
+ *  - number => Float64
+ *  - String => Utf8
+ *  - Buffer => Binary
+ *  - Record<String, any> => Struct
+ *  - Array<any> => List
+ *
  * @param data input data
  * @param options options to control the makeArrowTable call.
  *
@@ -86,8 +146,10 @@ export class MakeArrowTableOptions {
  *  ], { schema });
  * ```
  *
- * It guesses the vector columns if the schema is not provided. For example,
- * by default it assumes that the column named `vector` is a vector column.
+ * By default it assumes that the column named `vector` is a vector column
+ * and it will be converted into a fixed size list array of type float32.
+ * The `vectorColumns` option can be used to support other vector column
+ * names and data types.
  *
  * ```ts
  *
@@ -134,211 +196,304 @@ export function makeArrowTable (
   data: Array<Record<string, any>>,
   options?: Partial<MakeArrowTableOptions>
 ): ArrowTable {
-  if (data.length === 0) {
-    throw new Error('At least one record needs to be provided')
+  if (data.length === 0 && (options?.schema === undefined || options?.schema === null)) {
+    throw new Error('At least one record or a schema needs to be provided')
   }
+
   const opt = new MakeArrowTableOptions(options !== undefined ? options : {})
   const columns: Record<string, Vector> = {}
   // TODO: sample dataset to find missing columns
-  const columnNames = Object.keys(data[0])
+  // Prefer the field ordering of the schema, if present
+  const columnNames = ((options?.schema) != null) ? (options?.schema?.names as string[]) : Object.keys(data[0])
   for (const colName of columnNames) {
-    const values = data.map((datum) => datum[colName])
-    let vector: Vector
+    if (data.length !== 0 && !Object.prototype.hasOwnProperty.call(data[0], colName)) {
+      // The field is present in the schema, but not in the data, skip it
+      continue
+    }
+    // Extract a single column from the records (transpose from row-major to col-major)
+    let values = data.map((datum) => datum[colName])
 
+    // By default (type === undefined) arrow will infer the type from the JS type
+    let type
     if (opt.schema !== undefined) {
-      // Explicit schema is provided, highest priority
-      vector = vectorFromArray(
-        values,
-        opt.schema?.fields.filter((f) => f.name === colName)[0]?.type
-      )
+      // If there is a schema provided, then use that for the type instead
+      type = opt.schema?.fields.filter((f) => f.name === colName)[0]?.type
+      if (DataType.isInt(type) && type.bitWidth === 64) {
+        // wrap in BigInt to avoid bug: https://github.com/apache/arrow/issues/40051
+        values = values.map((v) => {
+          if (v === null) {
+            return v
+          }
+          return BigInt(v)
+        })
+      }
     } else {
+      // Otherwise, check to see if this column is one of the vector columns
+      // defined by opt.vectorColumns and, if so, use the fixed size list type
       const vectorColumnOptions = opt.vectorColumns[colName]
       if (vectorColumnOptions !== undefined) {
-        const fslType = new FixedSizeList(
-          values[0].length,
-          new Field('item', vectorColumnOptions.type, false)
-        )
-        vector = vectorFromArray(values, fslType)
-      } else {
-        // Normal case
-        vector = vectorFromArray(values)
+        type = newVectorType(values[0].length, vectorColumnOptions.type)
       }
     }
-    columns[colName] = vector
+
+    try {
+      // Convert an Array of JS values to an arrow vector
+      columns[colName] = makeVector(values, type, opt.dictionaryEncodeStrings)
+    } catch (error: unknown) {
+      // eslint-disable-next-line @typescript-eslint/restrict-template-expressions
+      throw Error(`Could not convert column "${colName}" to Arrow: ${error}`)
+    }
   }
 
-  return new ArrowTable(columns)
+  if (opt.schema != null) {
+    // `new ArrowTable(columns)` infers a schema which may sometimes have
+    // incorrect nullability (it assumes nullable=true if there are 0 rows)
+    //
+    // `new ArrowTable(schema, columns)` will also fail because it will create a
+    // batch with an inferred schema and then complain that the batch schema
+    // does not match the provided schema.
+    //
+    // To work around this we first create a table with the wrong schema and
+    // then patch the schema of the batches so we can use
+    // `new ArrowTable(schema, batches)` which does not do any schema inference
+    const firstTable = new ArrowTable(columns)
+    // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
+    const batchesFixed = firstTable.batches.map(batch => new RecordBatch(opt.schema!, batch.data))
+    return new ArrowTable(opt.schema, batchesFixed)
+  } else {
+    return new ArrowTable(columns)
+  }
 }
 
-// Converts an Array of records into an Arrow Table, optionally applying an embeddings function to it.
+/**
+ * Create an empty Arrow table with the provided schema
+ */
+export function makeEmptyTable (schema: Schema): ArrowTable {
+  return makeArrowTable([], { schema })
+}
+
+// Helper function to convert Array<Array<any>> to a variable sized list array
+function makeListVector (lists: any[][]): Vector<any> {
+  if (lists.length === 0 || lists[0].length === 0) {
+    throw Error('Cannot infer list vector from empty array or empty list')
+  }
+  const sampleList = lists[0]
+  let inferredType
+  try {
+    const sampleVector = makeVector(sampleList)
+    inferredType = sampleVector.type
+  } catch (error: unknown) {
+    // eslint-disable-next-line @typescript-eslint/restrict-template-expressions
+    throw Error(`Cannot infer list vector.  Cannot infer inner type: ${error}`)
+  }
+
+  const listBuilder = makeBuilder({
+    type: new List(new Field('item', inferredType, true))
+  })
+  for (const list of lists) {
+    listBuilder.append(list)
+  }
+  return listBuilder.finish().toVector()
+}
+
+// Helper function to convert an Array of JS values to an Arrow Vector
+function makeVector (values: any[], type?: DataType, stringAsDictionary?: boolean): Vector<any> {
+  if (type !== undefined) {
+    // No need for inference, let Arrow create it
+    return vectorFromArray(values, type)
+  }
+  if (values.length === 0) {
+    throw Error('makeVector requires at least one value or the type must be specfied')
+  }
+  const sampleValue = values.find(val => val !== null && val !== undefined)
+  if (sampleValue === undefined) {
+    throw Error('makeVector cannot infer the type if all values are null or undefined')
+  }
+  if (Array.isArray(sampleValue)) {
+    // Default Arrow inference doesn't handle list types
+    return makeListVector(values)
+  } else if (Buffer.isBuffer(sampleValue)) {
+    // Default Arrow inference doesn't handle Buffer
+    return vectorFromArray(values, new Binary())
+  } else if (!(stringAsDictionary ?? false) && (typeof sampleValue === 'string' || sampleValue instanceof String)) {
+    // If the type is string then don't use Arrow's default inference unless dictionaries are requested
+    // because it will always use dictionary encoding for strings
+    return vectorFromArray(values, new Utf8())
+  } else {
+    // Convert a JS array of values to an arrow vector
+    return vectorFromArray(values)
+  }
+}
+
+async function applyEmbeddings<T> (table: ArrowTable, embeddings?: EmbeddingFunction<T>, schema?: Schema): Promise<ArrowTable> {
+  if (embeddings == null) {
+    return table
+  }
+
+  // Convert from ArrowTable to Record<String, Vector>
+  const colEntries = [...Array(table.numCols).keys()].map((_, idx) => {
+    const name = table.schema.fields[idx].name
+    // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
+    const vec = table.getChildAt(idx)!
+    return [name, vec]
+  })
+  const newColumns = Object.fromEntries(colEntries)
+
+  const sourceColumn = newColumns[embeddings.sourceColumn]
+  const destColumn = embeddings.destColumn ?? 'vector'
+  const innerDestType = embeddings.embeddingDataType ?? new Float32()
+  if (sourceColumn === undefined) {
+    throw new Error(`Cannot apply embedding function because the source column '${embeddings.sourceColumn}' was not present in the data`)
+  }
+
+  if (table.numRows === 0) {
+    if (Object.prototype.hasOwnProperty.call(newColumns, destColumn)) {
+      // We have an empty table and it already has the embedding column so no work needs to be done
+      // Note: we don't return an error like we did below because this is a common occurrence.  For example,
+      // if we call convertToTable with 0 records and a schema that includes the embedding
+      return table
+    }
+    if (embeddings.embeddingDimension !== undefined) {
+      const destType = newVectorType(embeddings.embeddingDimension, innerDestType)
+      newColumns[destColumn] = makeVector([], destType)
+    } else if (schema != null) {
+      const destField = schema.fields.find(f => f.name === destColumn)
+      if (destField != null) {
+        newColumns[destColumn] = makeVector([], destField.type)
+      } else {
+        throw new Error(`Attempt to apply embeddings to an empty table failed because schema was missing embedding column '${destColumn}'`)
+      }
+    } else {
+      throw new Error('Attempt to apply embeddings to an empty table when the embeddings function does not specify `embeddingDimension`')
+    }
+  } else {
+    if (Object.prototype.hasOwnProperty.call(newColumns, destColumn)) {
+      throw new Error(`Attempt to apply embeddings to table failed because column ${destColumn} already existed`)
+    }
+    if (table.batches.length > 1) {
+      throw new Error('Internal error: `makeArrowTable` unexpectedly created a table with more than one batch')
+    }
+    const values = sourceColumn.toArray()
+    const vectors = await embeddings.embed(values as T[])
+    if (vectors.length !== values.length) {
+      throw new Error('Embedding function did not return an embedding for each input element')
+    }
+    const destType = newVectorType(vectors[0].length, innerDestType)
+    newColumns[destColumn] = makeVector(vectors, destType)
+  }
+
+  const newTable = new ArrowTable(newColumns)
+  if (schema != null) {
+    if (schema.fields.find(f => f.name === destColumn) === undefined) {
+      throw new Error(`When using embedding functions and specifying a schema the schema should include the embedding column but the column ${destColumn} was missing`)
+    }
+    return alignTable(newTable, schema)
+  }
+  return newTable
+}
+
+/*
+ * Convert an Array of records into an Arrow Table, optionally applying an
+ * embeddings function to it.
+ *
+ * This function calls `makeArrowTable` first to create the Arrow Table.
+ * Any provided `makeTableOptions` (e.g. a schema) will be passed on to
+ * that call.
+ *
+ * The embedding function will be passed a column of values (based on the
+ * `sourceColumn` of the embedding function) and expects to receive back
+ * number[][] which will be converted into a fixed size list column.  By
+ * default this will be a fixed size list of Float32 but that can be
+ * 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
+ * be placed at the end of the table, after all of the input columns.
+ */
 export async function convertToTable<T> (
   data: Array<Record<string, unknown>>,
-  embeddings?: EmbeddingFunction<T>
+  embeddings?: EmbeddingFunction<T>,
+  makeTableOptions?: Partial<MakeArrowTableOptions>
 ): Promise<ArrowTable> {
-  if (data.length === 0) {
-    throw new Error('At least one record needs to be provided')
-  }
-
-  const columns = Object.keys(data[0])
-  const records: Record<string, Vector> = {}
-
-  for (const columnsKey of columns) {
-    if (columnsKey === 'vector') {
-      const vectorSize = (data[0].vector as any[]).length
-      const listBuilder = newVectorBuilder(vectorSize)
-      for (const datum of data) {
-        if ((datum[columnsKey] as any[]).length !== vectorSize) {
-          throw new Error(`Invalid vector size, expected ${vectorSize}`)
-        }
-
-        listBuilder.append(datum[columnsKey])
-      }
-      records[columnsKey] = listBuilder.finish().toVector()
-    } else {
-      const values = []
-      for (const datum of data) {
-        values.push(datum[columnsKey])
-      }
-
-      if (columnsKey === embeddings?.sourceColumn) {
-        const vectors = await embeddings.embed(values as T[])
-        records.vector = vectorFromArray(
-          vectors,
-          newVectorType(vectors[0].length)
-        )
-      }
-
-      if (typeof values[0] === 'string') {
-        // `vectorFromArray` converts strings into dictionary vectors, forcing it back to a string column
-        records[columnsKey] = vectorFromArray(values, new Utf8())
-      } else if (Array.isArray(values[0])) {
-        const elementType = getElementType(values[0])
-        let innerType
-        if (elementType === 'string') {
-          innerType = new Utf8()
-        } else if (elementType === 'number') {
-          innerType = new Float64()
-        } else {
-          // TODO: pass in schema if it exists, else keep going to the next element
-          throw new Error(`Unsupported array element type ${elementType}`)
-        }
-        const listBuilder = makeBuilder({
-          type: new List(new Field('item', innerType, true))
-        })
-        for (const value of values) {
-          listBuilder.append(value)
-        }
-        records[columnsKey] = listBuilder.finish().toVector()
-      } else {
-        // TODO if this is a struct field then recursively align the subfields
-        records[columnsKey] = vectorFromArray(values)
-      }
-    }
-  }
-
-  return new ArrowTable(records)
-}
-
-function getElementType (arr: any[]): string {
-  if (arr.length === 0) {
-    return 'undefined'
-  }
-
-  return typeof arr[0]
-}
-
-// Creates a new Arrow ListBuilder that stores a Vector column
-function newVectorBuilder (dim: number): FixedSizeListBuilder<Float32> {
-  return makeBuilder({
-    type: newVectorType(dim)
-  })
+  const table = makeArrowTable(data, makeTableOptions)
+  return await applyEmbeddings(table, embeddings, makeTableOptions?.schema)
 }
 
 // Creates the Arrow Type for a Vector column with dimension `dim`
-function newVectorType (dim: number): FixedSizeList<Float32> {
+function newVectorType <T extends Float> (dim: number, innerType: T): FixedSizeList<T> {
   // Somewhere we always default to have the elements nullable, so we need to set it to true
   // otherwise we often get schema mismatches because the stored data always has schema with nullable elements
-  const children = new Field<Float32>('item', new Float32(), true)
+  const children = new Field<T>('item', innerType, true)
   return new FixedSizeList(dim, children)
 }
 
-// Converts an Array of records into Arrow IPC format
+/**
+ * Serialize an Array of records into a buffer using the Arrow IPC File serialization
+ *
+ * This function will call `convertToTable` and pass on `embeddings` and `schema`
+ *
+ * `schema` is required if data is empty
+ */
 export async function fromRecordsToBuffer<T> (
   data: Array<Record<string, unknown>>,
   embeddings?: EmbeddingFunction<T>,
   schema?: Schema
 ): Promise<Buffer> {
-  let table = await convertToTable(data, embeddings)
-  if (schema !== undefined) {
-    table = alignTable(table, schema)
-  }
+  const table = await convertToTable(data, embeddings, { schema })
   const writer = RecordBatchFileWriter.writeAll(table)
   return Buffer.from(await writer.toUint8Array())
 }
 
-// Converts an Array of records into Arrow IPC stream format
+/**
+ * Serialize an Array of records into a buffer using the Arrow IPC Stream serialization
+ *
+ * This function will call `convertToTable` and pass on `embeddings` and `schema`
+ *
+ * `schema` is required if data is empty
+ */
 export async function fromRecordsToStreamBuffer<T> (
   data: Array<Record<string, unknown>>,
   embeddings?: EmbeddingFunction<T>,
   schema?: Schema
 ): Promise<Buffer> {
-  let table = await convertToTable(data, embeddings)
-  if (schema !== undefined) {
-    table = alignTable(table, schema)
-  }
+  const table = await convertToTable(data, embeddings, { schema })
   const writer = RecordBatchStreamWriter.writeAll(table)
   return Buffer.from(await writer.toUint8Array())
 }
 
-// Converts an Arrow Table into Arrow IPC format
+/**
+ * Serialize an Arrow Table into a buffer using the Arrow IPC File serialization
+ *
+ * This function will apply `embeddings` to the table in a manner similar to
+ * `convertToTable`.
+ *
+ * `schema` is required if the table is empty
+ */
 export async function fromTableToBuffer<T> (
   table: ArrowTable,
   embeddings?: EmbeddingFunction<T>,
   schema?: Schema
 ): Promise<Buffer> {
-  if (embeddings !== undefined) {
-    const source = table.getChild(embeddings.sourceColumn)
-
-    if (source === null) {
-      throw new Error(
-        `The embedding source column ${embeddings.sourceColumn} was not found in the Arrow Table`
-      )
-    }
-
-    const vectors = await embeddings.embed(source.toArray() as T[])
-    const column = vectorFromArray(vectors, newVectorType(vectors[0].length))
-    table = table.assign(new ArrowTable({ vector: column }))
-  }
-  if (schema !== undefined) {
-    table = alignTable(table, schema)
-  }
-  const writer = RecordBatchFileWriter.writeAll(table)
+  const tableWithEmbeddings = await applyEmbeddings(table, embeddings, schema)
+  const writer = RecordBatchFileWriter.writeAll(tableWithEmbeddings)
   return Buffer.from(await writer.toUint8Array())
 }
 
-// Converts an Arrow Table into Arrow IPC stream format
+/**
+ * Serialize an Arrow Table into a buffer using the Arrow IPC Stream serialization
+ *
+ * This function will apply `embeddings` to the table in a manner similar to
+ * `convertToTable`.
+ *
+ * `schema` is required if the table is empty
+ */
 export async function fromTableToStreamBuffer<T> (
   table: ArrowTable,
   embeddings?: EmbeddingFunction<T>,
   schema?: Schema
 ): Promise<Buffer> {
-  if (embeddings !== undefined) {
-    const source = table.getChild(embeddings.sourceColumn)
-
-    if (source === null) {
-      throw new Error(
-        `The embedding source column ${embeddings.sourceColumn} was not found in the Arrow Table`
-      )
-    }
-
-    const vectors = await embeddings.embed(source.toArray() as T[])
-    const column = vectorFromArray(vectors, newVectorType(vectors[0].length))
-    table = table.assign(new ArrowTable({ vector: column }))
-  }
-  if (schema !== undefined) {
-    table = alignTable(table, schema)
-  }
-  const writer = RecordBatchStreamWriter.writeAll(table)
+  const tableWithEmbeddings = await applyEmbeddings(table, embeddings, schema)
+  const writer = RecordBatchStreamWriter.writeAll(tableWithEmbeddings)
   return Buffer.from(await writer.toUint8Array())
 }
 

node/src/embedding/embedding_function.ts

@@ -12,18 +12,53 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+import { type Float } from 'apache-arrow'
+
 /**
  * An embedding function that automatically creates vector representation for a given column.
  */
 export interface EmbeddingFunction<T> {
   /**
-     * The name of the column that will be used as input for the Embedding Function.
-     */
+   * The name of the column that will be used as input for the Embedding Function.
+   */
   sourceColumn: string
 
   /**
-     * Creates a vector representation for the given values.
-     */
+   * The data type of the embedding
+   *
+   * The embedding function should return `number`.  This will be converted into
+   * an Arrow float array.  By default this will be Float32 but this property can
+   * be used to control the conversion.
+   */
+  embeddingDataType?: Float
+
+  /**
+   * The dimension of the embedding
+   *
+   * This is optional, normally this can be determined by looking at the results of
+   * `embed`.  If this is not specified, and there is an attempt to apply the embedding
+   * to an empty table, then that process will fail.
+   */
+  embeddingDimension?: number
+
+  /**
+   * The name of the column that will contain the embedding
+   *
+   * By default this is "vector"
+   */
+  destColumn?: string
+
+  /**
+   * Should the source column be excluded from the resulting table
+   *
+   * By default the source column is included.  Set this to true and
+   * only the embedding will be stored.
+   */
+  excludeSource?: boolean
+
+  /**
+   * Creates a vector representation for the given values.
+   */
   embed: (data: T[]) => Promise<number[][]>
 }
 

node/src/index.ts

@@ -37,6 +37,7 @@ const {
   tableCountRows,
   tableDelete,
   tableUpdate,
+  tableMergeInsert,
   tableCleanupOldVersions,
   tableCompactFiles,
   tableListIndices,
@@ -48,7 +49,7 @@ const {
 export { Query }
 export type { EmbeddingFunction }
 export { OpenAIEmbeddingFunction } from './embedding/openai'
-export { makeArrowTable, type MakeArrowTableOptions } from './arrow'
+export { convertToTable, makeArrowTable, type MakeArrowTableOptions } from './arrow'
 
 const defaultAwsRegion = 'us-west-2'
 
@@ -163,6 +164,7 @@ export async function connect (
       {
         uri: '',
         awsCredentials: undefined,
+        awsRegion: defaultAwsRegion,
         apiKey: undefined,
         region: defaultAwsRegion
       },
@@ -174,7 +176,13 @@ export async function connect (
     // Remote connection
     return new RemoteConnection(opts)
   }
-  const db = await databaseNew(opts.uri)
+  const db = await databaseNew(
+    opts.uri,
+    opts.awsCredentials?.accessKeyId,
+    opts.awsCredentials?.secretKey,
+    opts.awsCredentials?.sessionToken,
+    opts.awsRegion
+  )
   return new LocalConnection(db, opts)
 }
 
@@ -364,7 +372,7 @@ export interface Table<T = number[]> {
   /**
    * Returns the number of rows in this table.
    */
-  countRows: () => Promise<number>
+  countRows: (filter?: string) => Promise<number>
 
   /**
    * Delete rows from this table.
@@ -433,6 +441,38 @@ export interface Table<T = number[]> {
    */
   update: (args: UpdateArgs | UpdateSqlArgs) => Promise<void>
 
+  /**
+   * Runs a "merge insert" operation on the table
+   *
+   * This operation can add rows, update rows, and remove rows all in a single
+   * transaction. It is a very generic tool that can be used to create
+   * behaviors like "insert if not exists", "update or insert (i.e. upsert)",
+   * or even replace a portion of existing data with new data (e.g. replace
+   * all data where month="january")
+   *
+   * The merge insert operation works by combining new data from a
+   * **source table** with existing data in a **target table** by using a
+   * join.  There are three categories of records.
+   *
+   * "Matched" records are records that exist in both the source table and
+   * the target table. "Not matched" records exist only in the source table
+   * (e.g. these are new data) "Not matched by source" records exist only
+   * in the target table (this is old data)
+   *
+   * The MergeInsertArgs can be used to customize what should happen for
+   * each category of data.
+   *
+   * Please note that the data may appear to be reordered as part of this
+   * operation.  This is because updated rows will be deleted from the
+   * dataset and then reinserted at the end with the new values.
+   *
+   * @param on a column to join on.  This is how records from the source
+   *           table and target table are matched.
+   * @param data the new data to insert
+   * @param args parameters controlling how the operation should behave
+   */
+  mergeInsert: (on: string, data: Array<Record<string, unknown>> | ArrowTable, args: MergeInsertArgs) => Promise<void>
+
   /**
    * List the indicies on this table.
    */
@@ -443,6 +483,8 @@ export interface Table<T = number[]> {
    */
   indexStats: (indexUuid: string) => Promise<IndexStats>
 
+  filter(value: string): Query<T>
+
   schema: Promise<Schema>
 }
 
@@ -474,6 +516,47 @@ export interface UpdateSqlArgs {
   valuesSql: Record<string, string>
 }
 
+export interface MergeInsertArgs {
+  /**
+   * If true then rows that exist in both the source table (new data) and
+   * the target table (old data) will be updated, replacing the old row
+   * with the corresponding matching row.
+   *
+   * If there are multiple matches then the behavior is undefined.
+   * Currently this causes multiple copies of the row to be created
+   * but that behavior is subject to change.
+   *
+   * Optionally, a filter can be specified.  This should be an SQL
+   * filter where fields with the prefix "target." refer to fields
+   * in the target table (old data) and fields with the prefix
+   * "source." refer to fields in the source table (new data).  For
+   * example, the filter "target.lastUpdated < source.lastUpdated" will
+   * only update matched rows when the incoming `lastUpdated` value is
+   * newer.
+   *
+   * Rows that do not match the filter will not be updated.  Rows that
+   * do not match the filter do become "not matched" rows.
+   */
+  whenMatchedUpdateAll?: string | boolean
+  /**
+   * If true then rows that exist only in the source table (new data)
+   * will be inserted into the target table.
+   */
+  whenNotMatchedInsertAll?: boolean
+  /**
+   * If true then rows that exist only in the target table (old data)
+   * will be deleted.
+   *
+   * If this is a string then it will be treated as an SQL filter and
+   * only rows that both do not match any row in the source table and
+   * match the given filter will be deleted.
+   *
+   * This can be used to replace a selection of existing data with
+   * new data.
+   */
+  whenNotMatchedBySourceDelete?: string | boolean
+}
+
 export interface VectorIndex {
   columns: string[]
   name: string
@@ -768,8 +851,8 @@ export class LocalTable<T = number[]> implements Table<T> {
   /**
    * Returns the number of rows in this table.
    */
-  async countRows (): Promise<number> {
-    return tableCountRows.call(this._tbl)
+  async countRows (filter?: string): Promise<number> {
+    return tableCountRows.call(this._tbl, filter)
   }
 
   /**
@@ -812,6 +895,46 @@ export class LocalTable<T = number[]> implements Table<T> {
       })
   }
 
+  async mergeInsert (on: string, data: Array<Record<string, unknown>> | ArrowTable, args: MergeInsertArgs): Promise<void> {
+    let whenMatchedUpdateAll = false
+    let whenMatchedUpdateAllFilt = null
+    if (args.whenMatchedUpdateAll !== undefined && args.whenMatchedUpdateAll !== null) {
+      whenMatchedUpdateAll = true
+      if (args.whenMatchedUpdateAll !== true) {
+        whenMatchedUpdateAllFilt = args.whenMatchedUpdateAll
+      }
+    }
+    const whenNotMatchedInsertAll = args.whenNotMatchedInsertAll ?? false
+    let whenNotMatchedBySourceDelete = false
+    let whenNotMatchedBySourceDeleteFilt = null
+    if (args.whenNotMatchedBySourceDelete !== undefined && args.whenNotMatchedBySourceDelete !== null) {
+      whenNotMatchedBySourceDelete = true
+      if (args.whenNotMatchedBySourceDelete !== true) {
+        whenNotMatchedBySourceDeleteFilt = args.whenNotMatchedBySourceDelete
+      }
+    }
+
+    const schema = await this.schema
+    let tbl: ArrowTable
+    if (data instanceof ArrowTable) {
+      tbl = data
+    } else {
+      tbl = makeArrowTable(data, { schema })
+    }
+    const buffer = await fromTableToBuffer(tbl, this._embeddings, schema)
+
+    this._tbl = await tableMergeInsert.call(
+      this._tbl,
+      on,
+      whenMatchedUpdateAll,
+      whenMatchedUpdateAllFilt,
+      whenNotMatchedInsertAll,
+      whenNotMatchedBySourceDelete,
+      whenNotMatchedBySourceDeleteFilt,
+      buffer
+    )
+  }
+
   /**
    * Clean up old versions of the table, freeing disk space.
    *

node/src/remote/index.ts

@@ -24,7 +24,8 @@ import {
   type IndexStats,
   type UpdateArgs,
   type UpdateSqlArgs,
-  makeArrowTable
+  makeArrowTable,
+  type MergeInsertArgs
 } from '../index'
 import { Query } from '../query'
 
@@ -270,6 +271,59 @@ export class RemoteTable<T = number[]> implements Table<T> {
     return new RemoteQuery(query, this._client, this._name) //, this._embeddings_new)
   }
 
+  filter (where: string): Query<T> {
+    throw new Error('Not implemented')
+  }
+
+  async mergeInsert (on: string, data: Array<Record<string, unknown>> | ArrowTable, args: MergeInsertArgs): Promise<void> {
+    let tbl: ArrowTable
+    if (data instanceof ArrowTable) {
+      tbl = data
+    } else {
+      tbl = makeArrowTable(data, await this.schema)
+    }
+
+    const queryParams: any = {
+      on
+    }
+    if (args.whenMatchedUpdateAll !== false && args.whenMatchedUpdateAll !== null && args.whenMatchedUpdateAll !== undefined) {
+      queryParams.when_matched_update_all = 'true'
+      if (typeof args.whenMatchedUpdateAll === 'string') {
+        queryParams.when_matched_update_all_filt = args.whenMatchedUpdateAll
+      }
+    } else {
+      queryParams.when_matched_update_all = 'false'
+    }
+    if (args.whenNotMatchedInsertAll ?? false) {
+      queryParams.when_not_matched_insert_all = 'true'
+    } else {
+      queryParams.when_not_matched_insert_all = 'false'
+    }
+    if (args.whenNotMatchedBySourceDelete !== false && args.whenNotMatchedBySourceDelete !== null && args.whenNotMatchedBySourceDelete !== undefined) {
+      queryParams.when_not_matched_by_source_delete = 'true'
+      if (typeof args.whenNotMatchedBySourceDelete === 'string') {
+        queryParams.when_not_matched_by_source_delete_filt = args.whenNotMatchedBySourceDelete
+      }
+    } else {
+      queryParams.when_not_matched_by_source_delete = 'false'
+    }
+
+    const buffer = await fromTableToStreamBuffer(tbl, this._embeddings)
+    const res = await this._client.post(
+      `/v1/table/${this._name}/merge_insert/`,
+      buffer,
+      queryParams,
+      'application/vnd.apache.arrow.stream'
+    )
+    if (res.status !== 200) {
+      throw new Error(
+        `Server Error, status: ${res.status}, ` +
+          // eslint-disable-next-line @typescript-eslint/restrict-template-expressions
+          `message: ${res.statusText}: ${res.data}`
+      )
+    }
+  }
+
   async add (data: Array<Record<string, unknown>> | ArrowTable): Promise<number> {
     let tbl: ArrowTable
     if (data instanceof ArrowTable) {

node/src/test/arrow.test.ts

@@ -13,9 +13,10 @@
 // limitations under the License.
 
 import { describe } from 'mocha'
-import { assert } from 'chai'
+import { assert, expect, use as chaiUse } from 'chai'
+import * as chaiAsPromised from 'chai-as-promised'
 
-import { fromTableToBuffer, makeArrowTable } from '../arrow'
+import { convertToTable, fromTableToBuffer, makeArrowTable, makeEmptyTable } from '../arrow'
 import {
   Field,
   FixedSizeList,
@@ -24,21 +25,79 @@ import {
   Int32,
   tableFromIPC,
   Schema,
-  Float64
+  Float64,
+  type Table,
+  Binary,
+  Bool,
+  Utf8,
+  Struct,
+  List,
+  DataType,
+  Dictionary,
+  Int64
 } from 'apache-arrow'
+import { type EmbeddingFunction } from '../embedding/embedding_function'
 
-describe('Apache Arrow tables', function () {
-  it('customized schema', async function () {
+chaiUse(chaiAsPromised)
+
+function sampleRecords (): Array<Record<string, any>> {
+  return [
+    {
+      binary: Buffer.alloc(5),
+      boolean: false,
+      number: 7,
+      string: 'hello',
+      struct: { x: 0, y: 0 },
+      list: ['anime', 'action', 'comedy']
+    }
+  ]
+}
+
+// Helper method to verify various ways to create a table
+async function checkTableCreation (tableCreationMethod: (records: any, recordsReversed: any, schema: Schema) => Promise<Table>): Promise<void> {
+  const records = sampleRecords()
+  const recordsReversed = [{
+    list: ['anime', 'action', 'comedy'],
+    struct: { x: 0, y: 0 },
+    string: 'hello',
+    number: 7,
+    boolean: false,
+    binary: Buffer.alloc(5)
+  }]
+  const schema = new Schema([
+    new Field('binary', new Binary(), false),
+    new Field('boolean', new Bool(), false),
+    new Field('number', new Float64(), false),
+    new Field('string', new Utf8(), false),
+    new Field('struct', new Struct([
+      new Field('x', new Float64(), false),
+      new Field('y', new Float64(), false)
+    ])),
+    new Field('list', new List(new Field('item', new Utf8(), false)), false)
+  ])
+
+  const table = await tableCreationMethod(records, recordsReversed, schema)
+  schema.fields.forEach((field, idx) => {
+    const actualField = table.schema.fields[idx]
+    assert.isFalse(actualField.nullable)
+    assert.equal(table.getChild(field.name)?.type.toString(), field.type.toString())
+    assert.equal(table.getChildAt(idx)?.type.toString(), field.type.toString())
+  })
+}
+
+describe('The function makeArrowTable', function () {
+  it('will use data types from a provided schema instead of inference', async function () {
     const schema = new Schema([
       new Field('a', new Int32()),
       new Field('b', new Float32()),
-      new Field('c', new FixedSizeList(3, new Field('item', new Float16())))
+      new Field('c', new FixedSizeList(3, new Field('item', new Float16()))),
+      new Field('d', new Int64())
     ])
     const table = makeArrowTable(
       [
-        { a: 1, b: 2, c: [1, 2, 3] },
-        { a: 4, b: 5, c: [4, 5, 6] },
-        { a: 7, b: 8, c: [7, 8, 9] }
+        { a: 1, b: 2, c: [1, 2, 3], d: 9 },
+        { a: 4, b: 5, c: [4, 5, 6], d: 10 },
+        { a: 7, b: 8, c: [7, 8, 9], d: null }
       ],
       { schema }
     )
@@ -52,13 +111,13 @@ describe('Apache Arrow tables', function () {
     assert.deepEqual(actualSchema, schema)
   })
 
-  it('default vector column', async function () {
+  it('will assume the column `vector` is FixedSizeList<Float32> by default', async function () {
     const schema = new Schema([
       new Field('a', new Float64()),
       new Field('b', new Float64()),
       new Field(
         'vector',
-        new FixedSizeList(3, new Field('item', new Float32()))
+        new FixedSizeList(3, new Field('item', new Float32(), true))
       )
     ])
     const table = makeArrowTable([
@@ -76,12 +135,12 @@ describe('Apache Arrow tables', function () {
     assert.deepEqual(actualSchema, schema)
   })
 
-  it('2 vector columns', async function () {
+  it('can support multiple vector columns', async function () {
     const schema = new Schema([
       new Field('a', new Float64()),
       new Field('b', new Float64()),
-      new Field('vec1', new FixedSizeList(3, new Field('item', new Float16()))),
-      new Field('vec2', new FixedSizeList(3, new Field('item', new Float16())))
+      new Field('vec1', new FixedSizeList(3, new Field('item', new Float16(), true))),
+      new Field('vec2', new FixedSizeList(3, new Field('item', new Float16(), true)))
     ])
     const table = makeArrowTable(
       [
@@ -105,4 +164,157 @@ describe('Apache Arrow tables', function () {
     const actualSchema = actual.schema
     assert.deepEqual(actualSchema, schema)
   })
+
+  it('will allow different vector column types', async function () {
+    const table = makeArrowTable(
+      [
+        { fp16: [1], fp32: [1], fp64: [1] }
+      ],
+      {
+        vectorColumns: {
+          fp16: { type: new Float16() },
+          fp32: { type: new Float32() },
+          fp64: { type: new Float64() }
+        }
+      }
+    )
+
+    assert.equal(table.getChild('fp16')?.type.children[0].type.toString(), new Float16().toString())
+    assert.equal(table.getChild('fp32')?.type.children[0].type.toString(), new Float32().toString())
+    assert.equal(table.getChild('fp64')?.type.children[0].type.toString(), new Float64().toString())
+  })
+
+  it('will use dictionary encoded strings if asked', async function () {
+    const table = makeArrowTable([{ str: 'hello' }])
+    assert.isTrue(DataType.isUtf8(table.getChild('str')?.type))
+
+    const tableWithDict = makeArrowTable([{ str: 'hello' }], { dictionaryEncodeStrings: true })
+    assert.isTrue(DataType.isDictionary(tableWithDict.getChild('str')?.type))
+
+    const schema = new Schema([
+      new Field('str', new Dictionary(new Utf8(), new Int32()))
+    ])
+
+    const tableWithDict2 = makeArrowTable([{ str: 'hello' }], { schema })
+    assert.isTrue(DataType.isDictionary(tableWithDict2.getChild('str')?.type))
+  })
+
+  it('will infer data types correctly', async function () {
+    await checkTableCreation(async (records) => makeArrowTable(records))
+  })
+
+  it('will allow a schema to be provided', async function () {
+    await checkTableCreation(async (records, _, schema) => makeArrowTable(records, { schema }))
+  })
+
+  it('will use the field order of any provided schema', async function () {
+    await checkTableCreation(async (_, recordsReversed, schema) => makeArrowTable(recordsReversed, { schema }))
+  })
+
+  it('will make an empty table', async function () {
+    await checkTableCreation(async (_, __, schema) => makeArrowTable([], { schema }))
+  })
+})
+
+class DummyEmbedding implements EmbeddingFunction<string> {
+  public readonly sourceColumn = 'string'
+  public readonly embeddingDimension = 2
+  public readonly embeddingDataType = new Float16()
+
+  async embed (data: string[]): Promise<number[][]> {
+    return data.map(
+      () => [0.0, 0.0]
+    )
+  }
+}
+
+class DummyEmbeddingWithNoDimension implements EmbeddingFunction<string> {
+  public readonly sourceColumn = 'string'
+
+  async embed (data: string[]): Promise<number[][]> {
+    return data.map(
+      () => [0.0, 0.0]
+    )
+  }
+}
+
+describe('convertToTable', function () {
+  it('will infer data types correctly', async function () {
+    await checkTableCreation(async (records) => await convertToTable(records))
+  })
+
+  it('will allow a schema to be provided', async function () {
+    await checkTableCreation(async (records, _, schema) => await convertToTable(records, undefined, { schema }))
+  })
+
+  it('will use the field order of any provided schema', async function () {
+    await checkTableCreation(async (_, recordsReversed, schema) => await convertToTable(recordsReversed, undefined, { schema }))
+  })
+
+  it('will make an empty table', async function () {
+    await checkTableCreation(async (_, __, schema) => await convertToTable([], undefined, { schema }))
+  })
+
+  it('will apply embeddings', async function () {
+    const records = sampleRecords()
+    const table = await convertToTable(records, new DummyEmbedding())
+    assert.isTrue(DataType.isFixedSizeList(table.getChild('vector')?.type))
+    assert.equal(table.getChild('vector')?.type.children[0].type.toString(), new Float16().toString())
+  })
+
+  it('will fail if missing the embedding source column', async function () {
+    return await expect(convertToTable([{ id: 1 }], new DummyEmbedding())).to.be.rejectedWith("'string' was not present")
+  })
+
+  it('use embeddingDimension if embedding missing from table', async function () {
+    const schema = new Schema([
+      new Field('string', new Utf8(), false)
+    ])
+    // Simulate getting an empty Arrow table (minus embedding) from some other source
+    // In other words, we aren't starting with records
+    const table = makeEmptyTable(schema)
+
+    // If the embedding specifies the dimension we are fine
+    await fromTableToBuffer(table, new DummyEmbedding())
+
+    // We can also supply a schema and should be ok
+    const schemaWithEmbedding = new Schema([
+      new Field('string', new Utf8(), false),
+      new Field('vector', new FixedSizeList(2, new Field('item', new Float16(), false)), false)
+    ])
+    await fromTableToBuffer(table, new DummyEmbeddingWithNoDimension(), schemaWithEmbedding)
+
+    // Otherwise we will get an error
+    return await expect(fromTableToBuffer(table, new DummyEmbeddingWithNoDimension())).to.be.rejectedWith('does not specify `embeddingDimension`')
+  })
+
+  it('will apply embeddings to an empty table', async function () {
+    const schema = new Schema([
+      new Field('string', new Utf8(), false),
+      new Field('vector', new FixedSizeList(2, new Field('item', new Float16(), false)), false)
+    ])
+    const table = await convertToTable([], new DummyEmbedding(), { schema })
+    assert.isTrue(DataType.isFixedSizeList(table.getChild('vector')?.type))
+    assert.equal(table.getChild('vector')?.type.children[0].type.toString(), new Float16().toString())
+  })
+
+  it('will complain if embeddings present but schema missing embedding column', async function () {
+    const schema = new Schema([
+      new Field('string', new Utf8(), false)
+    ])
+    return await expect(convertToTable([], new DummyEmbedding(), { schema })).to.be.rejectedWith('column vector was missing')
+  })
+
+  it('will provide a nice error if run twice', async function () {
+    const records = sampleRecords()
+    const table = await convertToTable(records, new DummyEmbedding())
+    // fromTableToBuffer will try and apply the embeddings again
+    return await expect(fromTableToBuffer(table, new DummyEmbedding())).to.be.rejectedWith('already existed')
+  })
+})
+
+describe('makeEmptyTable', function () {
+  it('will make an empty table', async function () {
+    await checkTableCreation(async (_, __, schema) => makeEmptyTable(schema))
+  })
 })

node/src/test/test.ts

@@ -294,6 +294,7 @@ describe('LanceDB client', function () {
       })
       assert.equal(table.name, 'vectors')
       assert.equal(await table.countRows(), 10)
+      assert.equal(await table.countRows('vector IS NULL'), 0)
       assert.deepEqual(await con.tableNames(), ['vectors'])
     })
 
@@ -369,6 +370,7 @@ describe('LanceDB client', function () {
       const table = await con.createTable('f16', data)
       assert.equal(table.name, 'f16')
       assert.equal(await table.countRows(), total)
+      assert.equal(await table.countRows('id < 5'), 5)
       assert.deepEqual(await con.tableNames(), ['f16'])
       assert.deepEqual(await table.schema, schema)
 
@@ -391,24 +393,6 @@ describe('LanceDB client', function () {
       })
     }).timeout(120000)
 
-    it('fails to create a new table when the vector column is missing', async function () {
-      const dir = await track().mkdir('lancejs')
-      const con = await lancedb.connect(dir)
-
-      const data = [
-        {
-          id: 1,
-          price: 10
-        }
-      ]
-
-      const create = con.createTable('missing_vector', data)
-      await expect(create).to.be.rejectedWith(
-        Error,
-        "column 'vector' is missing"
-      )
-    })
-
     it('use overwrite flag to overwrite existing table', async function () {
       const dir = await track().mkdir('lancejs')
       const con = await lancedb.connect(dir)
@@ -549,6 +533,54 @@ describe('LanceDB client', function () {
       assert.equal(await table.countRows(), 2)
     })
 
+    it('can merge insert records into the table', async function () {
+      const dir = await track().mkdir('lancejs')
+      const con = await lancedb.connect(dir)
+
+      const data = [{ id: 1, age: 1 }, { id: 2, age: 1 }]
+      const table = await con.createTable('my_table', data)
+
+      // insert if not exists
+      let newData = [{ id: 2, age: 2 }, { id: 3, age: 2 }]
+      await table.mergeInsert('id', newData, {
+        whenNotMatchedInsertAll: true
+      })
+      assert.equal(await table.countRows(), 3)
+      assert.equal(await table.countRows('age = 2'), 1)
+
+      // conditional update
+      newData = [{ id: 2, age: 3 }, { id: 3, age: 3 }]
+      await table.mergeInsert('id', newData, {
+        whenMatchedUpdateAll: 'target.age = 1'
+      })
+      assert.equal(await table.countRows(), 3)
+      assert.equal(await table.countRows('age = 1'), 1)
+      assert.equal(await table.countRows('age = 3'), 1)
+
+      newData = [{ id: 3, age: 4 }, { id: 4, age: 4 }]
+      await table.mergeInsert('id', newData, {
+        whenNotMatchedInsertAll: true,
+        whenMatchedUpdateAll: true
+      })
+      assert.equal(await table.countRows(), 4)
+      assert.equal((await table.filter('age = 4').execute()).length, 2)
+
+      newData = [{ id: 5, age: 5 }]
+      await table.mergeInsert('id', newData, {
+        whenNotMatchedInsertAll: true,
+        whenMatchedUpdateAll: true,
+        whenNotMatchedBySourceDelete: 'age < 4'
+      })
+      assert.equal(await table.countRows(), 3)
+
+      await table.mergeInsert('id', newData, {
+        whenNotMatchedInsertAll: true,
+        whenMatchedUpdateAll: true,
+        whenNotMatchedBySourceDelete: true
+      })
+      assert.equal(await table.countRows(), 1)
+    })
+
     it('can update records in the table', async function () {
       const uri = await createTestDB()
       const con = await lancedb.connect(uri)

node/tsconfig.json

@@ -9,6 +9,6 @@
     "declaration": true,
     "outDir": "./dist",
     "strict": true,
-    // "esModuleInterop": true,
+    "sourceMap": true,
   }
 }

nodejs/Cargo.toml

@@ -1,27 +1,30 @@
 [package]
 name = "vectordb-nodejs"
-edition = "2021"
+edition.workspace = true
 version = "0.0.0"
 license.workspace = true
+description.workspace = true
 repository.workspace = true
+keywords.workspace = true
+categories.workspace = true
 
 [lib]
 crate-type = ["cdylib"]
 
 [dependencies]
 arrow-ipc.workspace = true
-napi = { version = "2.14", default-features = false, features = [
+futures.workspace = true
+lance-linalg.workspace = true
+lance.workspace = true
+vectordb = { path = "../rust/vectordb" }
+napi = { version = "2.15", default-features = false, features = [
     "napi7",
     "async"
 ] }
-napi-derive = "2.14"
-vectordb = { path = "../rust/vectordb" }
-lance.workspace = true
-lance-linalg.workspace = true
+napi-derive = "2"
+
+# Prevent dynamic linking of lzma, which comes from datafusion
+lzma-sys = { version = "*", features = ["static"] }
 
 [build-dependencies]
 napi-build = "2.1"
-
-[profile.release]
-lto = true
-strip = "symbols"

nodejs/__test__/arrow.test.ts

@@ -14,6 +14,7 @@
 
 import { makeArrowTable, toBuffer } from "../vectordb/arrow";
 import {
+  Int64,
   Field,
   FixedSizeList,
   Float16,
@@ -104,3 +105,16 @@ test("2 vector columns", function () {
   const actualSchema = actual.schema;
   expect(actualSchema.toString()).toEqual(schema.toString());
 });
+
+test("handles int64", function() {
+  // https://github.com/lancedb/lancedb/issues/960
+  const schema = new Schema([
+    new Field("x", new Int64(), true)
+  ]);
+  const table = makeArrowTable([
+    { x: 1 },
+    { x: 2 },
+    { x: 3 }
+  ], { schema });
+  expect(table.schema).toEqual(schema);
+})

nodejs/__test__/table.test.ts

@@ -53,6 +53,16 @@ describe("Test creating index", () => {
     const indexDir = path.join(tmpDir, "test.lance", "_indices");
     expect(fs.readdirSync(indexDir)).toHaveLength(1);
     // TODO: check index type.
+
+    // Search without specifying the column
+    let query_vector = data.toArray()[5].vec.toJSON();
+    let rst = await tbl.query().nearestTo(query_vector).limit(2).toArrow();
+    expect(rst.numRows).toBe(2);
+
+    // Search with specifying the column
+    let rst2 = await tbl.search(query_vector, "vec").limit(2).toArrow();
+    expect(rst2.numRows).toBe(2);
+    expect(rst.toString()).toEqual(rst2.toString());
   });
 
   test("no vector column available", async () => {
@@ -71,6 +81,80 @@ describe("Test creating index", () => {
     await tbl.createIndex("val").build();
     const indexDir = path.join(tmpDir, "no_vec.lance", "_indices");
     expect(fs.readdirSync(indexDir)).toHaveLength(1);
+
+    for await (const r of tbl.query().filter("id > 1").select(["id"])) {
+      expect(r.numRows).toBe(1);
+    }
+  });
+
+  test("two columns with different dimensions", async () => {
+    const db = await connect(tmpDir);
+    const schema = new Schema([
+      new Field("id", new Int32(), true),
+      new Field("vec", new FixedSizeList(32, new Field("item", new Float32()))),
+      new Field(
+        "vec2",
+        new FixedSizeList(64, new Field("item", new Float32()))
+      ),
+    ]);
+    const tbl = await db.createTable(
+      "two_vectors",
+      makeArrowTable(
+        Array(300)
+          .fill(1)
+          .map((_, i) => ({
+            id: i,
+            vec: Array(32)
+              .fill(1)
+              .map(() => Math.random()),
+            vec2: Array(64) // different dimension
+              .fill(1)
+              .map(() => Math.random()),
+          })),
+        { schema }
+      )
+    );
+
+    // Only build index over v1
+    await expect(tbl.createIndex().build()).rejects.toThrow(
+      /.*More than one vector columns found.*/
+    );
+    tbl
+      .createIndex("vec")
+      .ivf_pq({ num_partitions: 2, num_sub_vectors: 2 })
+      .build();
+
+    const rst = await tbl
+      .query()
+      .nearestTo(
+        Array(32)
+          .fill(1)
+          .map(() => Math.random())
+      )
+      .limit(2)
+      .toArrow();
+    expect(rst.numRows).toBe(2);
+
+    // Search with specifying the column
+    await expect(
+      tbl
+        .search(
+          Array(64)
+            .fill(1)
+            .map(() => Math.random()),
+          "vec"
+        )
+        .limit(2)
+        .toArrow()
+    ).rejects.toThrow(/.*does not match the dimension.*/);
+
+    const query64 = Array(64)
+      .fill(1)
+      .map(() => Math.random());
+    const rst64_1 = await tbl.query().nearestTo(query64).limit(2).toArrow();
+    const rst64_2 = await tbl.search(query64, "vec2").limit(2).toArrow();
+    expect(rst64_1.toString()).toEqual(rst64_2.toString());
+    expect(rst64_1.numRows).toBe(2);
   });
 
   test("create scalar index", async () => {

nodejs/jest.config.js

@@ -2,4 +2,6 @@
 module.exports = {
   preset: 'ts-jest',
   testEnvironment: 'node',
-};
+  moduleDirectories: ["node_modules", "./dist"],
+  moduleFileExtensions: ["js", "ts"],
+};

nodejs/src/index.rs

@@ -91,7 +91,6 @@ impl IndexBuilder {
 
     #[napi]
     pub async fn build(&self) -> napi::Result<()> {
-        println!("nodejs::index.rs : build");
         self.inner
             .build()
             .await

nodejs/src/iterator.rs

@@ -0,0 +1,47 @@
+// Copyright 2024 Lance Developers.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use futures::StreamExt;
+use lance::io::RecordBatchStream;
+use napi::bindgen_prelude::*;
+use napi_derive::napi;
+use vectordb::ipc::batches_to_ipc_file;
+
+/** Typescript-style Async Iterator over RecordBatches */
+#[napi]
+pub struct RecordBatchIterator {
+    inner: Box<dyn RecordBatchStream + Unpin>,
+}
+
+#[napi]
+impl RecordBatchIterator {
+    pub(crate) fn new(inner: Box<dyn RecordBatchStream + Unpin>) -> Self {
+        Self { inner }
+    }
+
+    #[napi]
+    pub async unsafe fn next(&mut self) -> napi::Result<Option<Buffer>> {
+        if let Some(rst) = self.inner.next().await {
+            let batch = rst.map_err(|e| {
+                napi::Error::from_reason(format!("Failed to get next batch from stream: {}", e))
+            })?;
+            batches_to_ipc_file(&[batch])
+                .map_err(|e| napi::Error::from_reason(format!("Failed to write IPC file: {}", e)))
+                .map(|buf| Some(Buffer::from(buf)))
+        } else {
+            // We are done with the stream.
+            Ok(None)
+        }
+    }
+}

nodejs/src/lib.rs

@@ -17,6 +17,7 @@ use napi_derive::*;
 
 mod connection;
 mod index;
+mod iterator;
 mod query;
 mod table;
 

nodejs/src/query.rs

@@ -16,7 +16,7 @@ use napi::bindgen_prelude::*;
 use napi_derive::napi;
 use vectordb::query::Query as LanceDBQuery;
 
-use crate::table::Table;
+use crate::{iterator::RecordBatchIterator, table::Table};
 
 #[napi]
 pub struct Query {
@@ -32,17 +32,50 @@ impl Query {
     }
 
     #[napi]
-    pub fn vector(&mut self, vector: Float32Array) {
-        let inn = self.inner.clone().nearest_to(&vector);
-        self.inner = inn;
+    pub fn column(&mut self, column: String) {
+        self.inner = self.inner.clone().column(&column);
     }
 
     #[napi]
-    pub fn to_arrow(&self) -> napi::Result<()> {
-        // let buf = self.inner.to_arrow().map_err(|e| {
-        //     napi::Error::from_reason(format!("Failed to convert query to arrow: {}", e))
-        // })?;
-        // Ok(buf)
-        todo!()
+    pub fn filter(&mut self, filter: String) {
+        self.inner = self.inner.clone().filter(filter);
+    }
+
+    #[napi]
+    pub fn select(&mut self, columns: Vec<String>) {
+        self.inner = self.inner.clone().select(&columns);
+    }
+
+    #[napi]
+    pub fn limit(&mut self, limit: u32) {
+        self.inner = self.inner.clone().limit(limit as usize);
+    }
+
+    #[napi]
+    pub fn prefilter(&mut self, prefilter: bool) {
+        self.inner = self.inner.clone().prefilter(prefilter);
+    }
+
+    #[napi]
+    pub fn nearest_to(&mut self, vector: Float32Array) {
+        self.inner = self.inner.clone().nearest_to(&vector);
+    }
+
+    #[napi]
+    pub fn refine_factor(&mut self, refine_factor: u32) {
+        self.inner = self.inner.clone().refine_factor(refine_factor);
+    }
+
+    #[napi]
+    pub fn nprobes(&mut self, nprobe: u32) {
+        self.inner = self.inner.clone().nprobes(nprobe as usize);
+    }
+
+    #[napi]
+    pub async fn execute_stream(&self) -> napi::Result<RecordBatchIterator> {
+        let inner_stream = self.inner.execute_stream().await.map_err(|e| {
+            napi::Error::from_reason(format!("Failed to execute query stream: {}", e))
+        })?;
+        Ok(RecordBatchIterator::new(Box::new(inner_stream)))
     }
 }

nodejs/src/table.rs

@@ -57,8 +57,8 @@ impl Table {
     }
 
     #[napi]
-    pub async fn count_rows(&self) -> napi::Result<usize> {
-        self.table.count_rows().await.map_err(|e| {
+    pub async fn count_rows(&self, filter: Option<String>) -> napi::Result<usize> {
+        self.table.count_rows(filter).await.map_err(|e| {
             napi::Error::from_reason(format!(
                 "Failed to count rows in table {}: {}",
                 self.table, e

nodejs/vectordb/arrow.ts

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 import {
+  Int64,
   Field,
   FixedSizeList,
   Float,
@@ -23,6 +24,7 @@ import {
   Vector,
   vectorFromArray,
   tableToIPC,
+  DataType,
 } from "apache-arrow";
 
 /** Data type accepted by NodeJS SDK */
@@ -137,15 +139,18 @@ export function makeArrowTable(
   const columnNames = Object.keys(data[0]);
   for (const colName of columnNames) {
     // eslint-disable-next-line @typescript-eslint/no-unsafe-return
-    const values = data.map((datum) => datum[colName]);
+    let values = data.map((datum) => datum[colName]);
     let vector: Vector;
 
     if (opt.schema !== undefined) {
       // Explicit schema is provided, highest priority
-      vector = vectorFromArray(
-        values,
-        opt.schema?.fields.filter((f) => f.name === colName)[0]?.type
-      );
+      const fieldType: DataType | undefined = opt.schema.fields.filter((f) => f.name === colName)[0]?.type as DataType;
+      if (fieldType instanceof Int64) {
+        // wrap in BigInt to avoid bug: https://github.com/apache/arrow/issues/40051
+        // eslint-disable-next-line @typescript-eslint/no-unsafe-argument
+        values = values.map((v) => BigInt(v));
+      }
+      vector = vectorFromArray(values, fieldType);
     } else {
       const vectorColumnOptions = opt.vectorColumns[colName];
       if (vectorColumnOptions !== undefined) {

nodejs/vectordb/native.d.ts

@@ -54,15 +54,26 @@ export class IndexBuilder {
   scalar(): void
   build(): Promise<void>
 }
+/** Typescript-style Async Iterator over RecordBatches  */
+export class RecordBatchIterator {
+  next(): Promise<Buffer | null>
+}
 export class Query {
-  vector(vector: Float32Array): void
-  toArrow(): void
+  column(column: string): void
+  filter(filter: string): void
+  select(columns: Array<string>): void
+  limit(limit: number): void
+  prefilter(prefilter: boolean): void
+  nearestTo(vector: Float32Array): void
+  refineFactor(refineFactor: number): void
+  nprobes(nprobe: number): void
+  executeStream(): Promise<RecordBatchIterator>
 }
 export class Table {
   /** Return Schema as empty Arrow IPC file. */
   schema(): Buffer
   add(buf: Buffer): Promise<void>
-  countRows(): Promise<bigint>
+  countRows(filter?: string | undefined | null): Promise<bigint>
   delete(predicate: string): Promise<void>
   createIndex(): IndexBuilder
   query(): Query

nodejs/vectordb/native.js

@@ -295,12 +295,13 @@ if (!nativeBinding) {
   throw new Error(`Failed to load native binding`)
 }
 
-const { Connection, IndexType, MetricType, IndexBuilder, Query, Table, WriteMode, connect } = nativeBinding
+const { Connection, IndexType, MetricType, IndexBuilder, RecordBatchIterator, Query, Table, WriteMode, connect } = nativeBinding
 
 module.exports.Connection = Connection
 module.exports.IndexType = IndexType
 module.exports.MetricType = MetricType
 module.exports.IndexBuilder = IndexBuilder
+module.exports.RecordBatchIterator = RecordBatchIterator
 module.exports.Query = Query
 module.exports.Table = Table
 module.exports.WriteMode = WriteMode

nodejs/vectordb/query.ts

@@ -12,46 +12,73 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-import { RecordBatch } from "apache-arrow";
-import { Table } from "./table";
+import { RecordBatch, tableFromIPC, Table as ArrowTable } from "apache-arrow";
+import {
+  RecordBatchIterator as NativeBatchIterator,
+  Query as NativeQuery,
+  Table as NativeTable,
+} from "./native";
 
-// TODO: re-eanble eslint once we have a real implementation
-/* eslint-disable */
 class RecordBatchIterator implements AsyncIterator<RecordBatch> {
-  next(
-    ...args: [] | [undefined]
-  ): Promise<IteratorResult<RecordBatch<any>, any>> {
-    throw new Error("Method not implemented.");
+  private promised_inner?: Promise<NativeBatchIterator>;
+  private inner?: NativeBatchIterator;
+
+  constructor(
+    inner?: NativeBatchIterator,
+    promise?: Promise<NativeBatchIterator>
+  ) {
+    // TODO: check promise reliably so we dont need to pass two arguments.
+    this.inner = inner;
+    this.promised_inner = promise;
   }
-  return?(value?: any): Promise<IteratorResult<RecordBatch<any>, any>> {
-    throw new Error("Method not implemented.");
-  }
-  throw?(e?: any): Promise<IteratorResult<RecordBatch<any>, any>> {
-    throw new Error("Method not implemented.");
+
+  async next(): Promise<IteratorResult<RecordBatch<any>, any>> {
+    if (this.inner === undefined) {
+      this.inner = await this.promised_inner;
+    }
+    if (this.inner === undefined) {
+      throw new Error("Invalid iterator state state");
+    }
+    const n = await this.inner.next();
+    if (n == null) {
+      return Promise.resolve({ done: true, value: null });
+    }
+    const tbl = tableFromIPC(n);
+    if (tbl.batches.length != 1) {
+      throw new Error("Expected only one batch");
+    }
+    return Promise.resolve({ done: false, value: tbl.batches[0] });
   }
 }
 /* eslint-enable */
 
 /** Query executor */
 export class Query implements AsyncIterable<RecordBatch> {
-  private readonly tbl: Table;
-  private _filter?: string;
-  private _limit?: number;
+  private readonly inner: NativeQuery;
 
-  // Vector search
-  private _vector?: Float32Array;
-  private _nprobes?: number;
-  private _refine_factor?: number = 1;
+  constructor(tbl: NativeTable) {
+    this.inner = tbl.query();
+  }
 
-  constructor(tbl: Table) {
-    this.tbl = tbl;
+  /** Set the column to run query. */
+  column(column: string): Query {
+    this.inner.column(column);
+    return this;
   }
 
   /** Set the filter predicate, only returns the results that satisfy the filter.
    *
    */
   filter(predicate: string): Query {
-    this._filter = predicate;
+    this.inner.filter(predicate);
+    return this;
+  }
+
+  /**
+   * Select the columns to return. If not set, all columns are returned.
+   */
+  select(columns: string[]): Query {
+    this.inner.select(columns);
     return this;
   }
 
@@ -59,35 +86,67 @@ export class Query implements AsyncIterable<RecordBatch> {
    * Set the limit of rows to return.
    */
   limit(limit: number): Query {
-    this._limit = limit;
+    this.inner.limit(limit);
+    return this;
+  }
+
+  prefilter(prefilter: boolean): Query {
+    this.inner.prefilter(prefilter);
     return this;
   }
 
   /**
    * Set the query vector.
    */
-  vector(vector: number[]): Query {
-    this._vector = Float32Array.from(vector);
+  nearestTo(vector: number[]): Query {
+    this.inner.nearestTo(Float32Array.from(vector));
     return this;
   }
 
   /**
-   * Set the number of probes to use for the query.
+   * Set the number of IVF partitions to use for the query.
    */
   nprobes(nprobes: number): Query {
-    this._nprobes = nprobes;
+    this.inner.nprobes(nprobes);
     return this;
   }
 
   /**
    * Set the refine factor for the query.
    */
-  refine_factor(refine_factor: number): Query {
-    this._refine_factor = refine_factor;
+  refineFactor(refine_factor: number): Query {
+    this.inner.refineFactor(refine_factor);
     return this;
   }
 
-  [Symbol.asyncIterator](): AsyncIterator<RecordBatch<any>, any, undefined> {
-    throw new RecordBatchIterator();
+  /**
+   * Execute the query and return the results as an AsyncIterator.
+   */
+  async executeStream(): Promise<RecordBatchIterator> {
+    const inner = await this.inner.executeStream();
+    return new RecordBatchIterator(inner);
+  }
+
+  /** Collect the results as an Arrow Table. */
+  async toArrow(): Promise<ArrowTable> {
+    const batches = [];
+    for await (const batch of this) {
+      batches.push(batch);
+    }
+    return new ArrowTable(batches);
+  }
+
+  /** Returns a JSON Array of All results.
+   *
+   */
+  async toArray(): Promise<any[]> {
+    const tbl = await this.toArrow();
+    // eslint-disable-next-line @typescript-eslint/no-unsafe-return
+    return tbl.toArray();
+  }
+
+  [Symbol.asyncIterator](): AsyncIterator<RecordBatch<any>> {
+    const promise = this.inner.executeStream();
+    return new RecordBatchIterator(undefined, promise);
   }
 }

nodejs/vectordb/table.ts

@@ -50,8 +50,8 @@ export class Table {
   }
 
   /** Count the total number of rows in the dataset. */
-  async countRows(): Promise<bigint> {
-    return await this.inner.countRows();
+  async countRows(filter?: string): Promise<bigint> {
+    return await this.inner.countRows(filter);
   }
 
   /** Delete the rows that satisfy the predicate. */
@@ -95,10 +95,58 @@ export class Table {
     return builder;
   }
 
-  search(vector?: number[]): Query {
-    const q = new Query(this);
-    if (vector !== undefined) {
-      q.vector(vector);
+  /**
+   * Create a generic {@link Query} Builder.
+   *
+   * When appropriate, various indices and statistics based pruning will be used to
+   * accelerate the query.
+   *
+   * @example
+   *
+   * ### Run a SQL-style query
+   * ```typescript
+   * for await (const batch of table.query()
+   *                          .filter("id > 1").select(["id"]).limit(20)) {
+   *  console.log(batch);
+   * }
+   * ```
+   *
+   * ### Run Top-10 vector similarity search
+   * ```typescript
+   * for await (const batch of table.query()
+   *                    .nearestTo([1, 2, 3])
+   *                    .refineFactor(5).nprobe(10)
+   *                    .limit(10)) {
+   *  console.log(batch);
+   * }
+   *```
+   *
+   * ### Scan the full dataset
+   * ```typescript
+   * for await (const batch of table.query()) {
+   *   console.log(batch);
+   * }
+   *
+   * ### Return the full dataset as Arrow Table
+   * ```typescript
+   * let arrowTbl = await table.query().nearestTo([1.0, 2.0, 0.5, 6.7]).toArrow();
+   * ```
+   *
+   * @returns {@link Query}
+   */
+  query(): Query {
+    return new Query(this.inner);
+  }
+
+  /** Search the table with a given query vector.
+   *
+   * This is a convenience method for preparing an ANN {@link Query}.
+   */
+  search(vector: number[], column?: string): Query {
+    const q = this.query();
+    q.nearestTo(vector);
+    if (column !== undefined) {
+      q.column(column);
     }
     return q;
   }

python/.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.5.1
+current_version = 0.5.5
 commit = True
 message = [python] Bump version: {current_version} → {new_version}
 tag = True

python/README.md

@@ -42,6 +42,12 @@ To run the unit tests:
 pytest
 ```
 
+To run the doc tests:
+
+```bash
+pytest --doctest-modules lancedb
+```
+
 To run linter and automatically fix all errors:
 
 ```bash

python/lancedb/__init__.py

@@ -13,6 +13,7 @@
 
 import importlib.metadata
 import os
+from datetime import timedelta
 from typing import Optional
 
 __version__ = importlib.metadata.version("lancedb")
@@ -30,6 +31,7 @@ def connect(
     api_key: Optional[str] = None,
     region: str = "us-east-1",
     host_override: Optional[str] = None,
+    read_consistency_interval: Optional[timedelta] = None,
 ) -> DBConnection:
     """Connect to a LanceDB database.
 
@@ -45,6 +47,18 @@ def connect(
         The region to use for LanceDB Cloud.
     host_override: str, optional
         The override url for LanceDB Cloud.
+    read_consistency_interval: timedelta, default None
+        (For LanceDB OSS only)
+        The interval at which to check for updates to the table from other
+        processes. If None, then consistency is not checked. For performance
+        reasons, this is the default. For strong consistency, set this to
+        zero seconds. Then every read will check for updates from other
+        processes. As a compromise, you can set this to a non-zero timedelta
+        for eventual consistency. If more than that interval has passed since
+        the last check, then the table will be checked for updates. Note: this
+        consistency only applies to read operations. Write operations are
+        always consistent.
+
 
     Examples
     --------
@@ -73,4 +87,4 @@ def connect(
         if api_key is None:
             raise ValueError(f"api_key is required to connected LanceDB cloud: {uri}")
         return RemoteDBConnection(uri, api_key, region, host_override)
-    return LanceDBConnection(uri)
+    return LanceDBConnection(uri, read_consistency_interval=read_consistency_interval)

python/lancedb/db.py

@@ -26,6 +26,8 @@ from .table import LanceTable, Table
 from .util import fs_from_uri, get_uri_location, get_uri_scheme, join_uri
 
 if TYPE_CHECKING:
+    from datetime import timedelta
+
     from .common import DATA, URI
     from .embeddings import EmbeddingFunctionConfig
     from .pydantic import LanceModel
@@ -118,7 +120,7 @@ class DBConnection(EnforceOverrides):
         >>> data = [{"vector": [1.1, 1.2], "lat": 45.5, "long": -122.7},
         ...         {"vector": [0.2, 1.8], "lat": 40.1, "long":  -74.1}]
         >>> db.create_table("my_table", data)
-        LanceTable(my_table)
+        LanceTable(connection=..., name="my_table")
         >>> db["my_table"].head()
         pyarrow.Table
         vector: fixed_size_list<item: float>[2]
@@ -139,7 +141,7 @@ class DBConnection(EnforceOverrides):
         ...    "long": [-122.7, -74.1]
         ... })
         >>> db.create_table("table2", data)
-        LanceTable(table2)
+        LanceTable(connection=..., name="table2")
         >>> db["table2"].head()
         pyarrow.Table
         vector: fixed_size_list<item: float>[2]
@@ -161,7 +163,7 @@ class DBConnection(EnforceOverrides):
         ...   pa.field("long", pa.float32())
         ... ])
         >>> db.create_table("table3", data, schema = custom_schema)
-        LanceTable(table3)
+        LanceTable(connection=..., name="table3")
         >>> db["table3"].head()
         pyarrow.Table
         vector: fixed_size_list<item: float>[2]
@@ -195,7 +197,7 @@ class DBConnection(EnforceOverrides):
         ...     pa.field("price", pa.float32()),
         ... ])
         >>> db.create_table("table4", make_batches(), schema=schema)
-        LanceTable(table4)
+        LanceTable(connection=..., name="table4")
 
         """
         raise NotImplementedError
@@ -243,6 +245,16 @@ class LanceDBConnection(DBConnection):
     ----------
     uri: str or Path
         The root uri of the database.
+    read_consistency_interval: timedelta, default None
+        The interval at which to check for updates to the table from other
+        processes. If None, then consistency is not checked. For performance
+        reasons, this is the default. For strong consistency, set this to
+        zero seconds. Then every read will check for updates from other
+        processes. As a compromise, you can set this to a non-zero timedelta
+        for eventual consistency. If more than that interval has passed since
+        the last check, then the table will be checked for updates. Note: this
+        consistency only applies to read operations. Write operations are
+        always consistent.
 
     Examples
     --------
@@ -250,22 +262,24 @@ class LanceDBConnection(DBConnection):
     >>> db = lancedb.connect("./.lancedb")
     >>> db.create_table("my_table", data=[{"vector": [1.1, 1.2], "b": 2},
     ...                                   {"vector": [0.5, 1.3], "b": 4}])
-    LanceTable(my_table)
+    LanceTable(connection=..., name="my_table")
     >>> db.create_table("another_table", data=[{"vector": [0.4, 0.4], "b": 6}])
-    LanceTable(another_table)
+    LanceTable(connection=..., name="another_table")
     >>> sorted(db.table_names())
     ['another_table', 'my_table']
     >>> len(db)
     2
     >>> db["my_table"]
-    LanceTable(my_table)
+    LanceTable(connection=..., name="my_table")
     >>> "my_table" in db
     True
     >>> db.drop_table("my_table")
     >>> db.drop_table("another_table")
     """
 
-    def __init__(self, uri: URI):
+    def __init__(
+        self, uri: URI, *, read_consistency_interval: Optional[timedelta] = None
+    ):
         if not isinstance(uri, Path):
             scheme = get_uri_scheme(uri)
         is_local = isinstance(uri, Path) or scheme == "file"
@@ -277,6 +291,14 @@ class LanceDBConnection(DBConnection):
         self._uri = str(uri)
 
         self._entered = False
+        self.read_consistency_interval = read_consistency_interval
+
+    def __repr__(self) -> str:
+        val = f"{self.__class__.__name__}({self._uri}"
+        if self.read_consistency_interval is not None:
+            val += f", read_consistency_interval={repr(self.read_consistency_interval)}"
+        val += ")"
+        return val
 
     @property
     def uri(self) -> str:

python/lancedb/embeddings/__init__.py

@@ -13,6 +13,7 @@
 
 # ruff: noqa: F401
 from .base import EmbeddingFunction, EmbeddingFunctionConfig, TextEmbeddingFunction
+from .bedrock import BedRockText
 from .cohere import CohereEmbeddingFunction
 from .gemini_text import GeminiText
 from .instructor import InstructorEmbeddingFunction

python/lancedb/embeddings/base.py

@@ -10,7 +10,6 @@
 #  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 #  See the License for the specific language governing permissions and
 #  limitations under the License.
-import importlib
 from abc import ABC, abstractmethod
 from typing import List, Union
 
@@ -91,25 +90,6 @@ class EmbeddingFunction(BaseModel, ABC):
             texts = texts.combine_chunks().to_pylist()
         return texts
 
-    @classmethod
-    def safe_import(cls, module: str, mitigation=None):
-        """
-        Import the specified module. If the module is not installed,
-        raise an ImportError with a helpful message.
-
-        Parameters
-        ----------
-        module : str
-            The name of the module to import
-        mitigation : Optional[str]
-            The package(s) to install to mitigate the error.
-            If not provided then the module name will be used.
-        """
-        try:
-            return importlib.import_module(module)
-        except ImportError:
-            raise ImportError(f"Please install {mitigation or module}")
-
     def safe_model_dump(self):
         from ..pydantic import PYDANTIC_VERSION
 

python/lancedb/embeddings/bedrock.py

@@ -0,0 +1,224 @@
+#  Copyright (c) 2023. LanceDB Developers
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+
+import json
+from functools import cached_property
+from typing import List, Union
+
+import numpy as np
+
+from lancedb.pydantic import PYDANTIC_VERSION
+
+from ..util import attempt_import_or_raise
+from .base import TextEmbeddingFunction
+from .registry import register
+from .utils import TEXT
+
+
+@register("bedrock-text")
+class BedRockText(TextEmbeddingFunction):
+    """
+    Parameters
+    ----------
+    name: str, default "amazon.titan-embed-text-v1"
+        The model ID of the bedrock model to use. Supported models for are:
+        - amazon.titan-embed-text-v1
+        - cohere.embed-english-v3
+        - cohere.embed-multilingual-v3
+    region: str, default "us-east-1"
+        Optional name of the AWS Region in which the service should be called.
+    profile_name: str, default None
+        Optional name of the AWS profile to use for calling the Bedrock service.
+        If not specified, the default profile will be used.
+    assumed_role: str, default None
+        Optional ARN of an AWS IAM role to assume for calling the Bedrock service.
+        If not specified, the current active credentials will be used.
+    role_session_name: str, default "lancedb-embeddings"
+        Optional name of the AWS IAM role session to use for calling the Bedrock
+        service. If not specified, "lancedb-embeddings" name will be used.
+
+    Examples
+    --------
+    import lancedb
+    import pandas as pd
+    from lancedb.pydantic import LanceModel, Vector
+
+    model = get_registry().get("bedrock-text").create()
+
+    class TextModel(LanceModel):
+        text: str = model.SourceField()
+        vector: Vector(model.ndims()) = model.VectorField()
+
+    df = pd.DataFrame({"text": ["hello world", "goodbye world"]})
+    db = lancedb.connect("tmp_path")
+    tbl = db.create_table("test", schema=TextModel, mode="overwrite")
+
+    tbl.add(df)
+
+    rs = tbl.search("hello").limit(1).to_pandas()
+    """
+
+    name: str = "amazon.titan-embed-text-v1"
+    region: str = "us-east-1"
+    assumed_role: Union[str, None] = None
+    profile_name: Union[str, None] = None
+    role_session_name: str = "lancedb-embeddings"
+
+    if PYDANTIC_VERSION < (2, 0):  # Pydantic 1.x compat
+
+        class Config:
+            keep_untouched = (cached_property,)
+
+    def ndims(self):
+        # return len(self._generate_embedding("test"))
+        # TODO: fix hardcoding
+        if self.name == "amazon.titan-embed-text-v1":
+            return 1536
+        elif self.name in {"cohere.embed-english-v3", "cohere.embed-multilingual-v3"}:
+            return 1024
+        else:
+            raise ValueError(f"Unknown model name: {self.name}")
+
+    def compute_query_embeddings(
+        self, query: str, *args, **kwargs
+    ) -> List[List[float]]:
+        return self.compute_source_embeddings(query)
+
+    def compute_source_embeddings(
+        self, texts: TEXT, *args, **kwargs
+    ) -> List[List[float]]:
+        texts = self.sanitize_input(texts)
+        return self.generate_embeddings(texts)
+
+    def generate_embeddings(
+        self, texts: Union[List[str], np.ndarray], *args, **kwargs
+    ) -> List[List[float]]:
+        """
+        Get the embeddings for the given texts
+
+        Parameters
+        ----------
+        texts: list[str] or np.ndarray (of str)
+            The texts to embed
+
+        Returns
+        -------
+        list[list[float]]
+            The embeddings for the given texts
+        """
+        results = []
+        for text in texts:
+            response = self._generate_embedding(text)
+            results.append(response)
+        return results
+
+    def _generate_embedding(self, text: str) -> List[float]:
+        """
+        Get the embeddings for the given texts
+
+        Parameters
+        ----------
+        texts: str
+            The texts to embed
+
+        Returns
+        -------
+        list[float]
+            The embeddings for the given texts
+        """
+        # format input body for provider
+        provider = self.name.split(".")[0]
+        _model_kwargs = {}
+        input_body = {**_model_kwargs}
+        if provider == "cohere":
+            if "input_type" not in input_body.keys():
+                input_body["input_type"] = "search_document"
+            input_body["texts"] = [text]
+        else:
+            # includes common provider == "amazon"
+            input_body["inputText"] = text
+        body = json.dumps(input_body)
+
+        try:
+            # invoke bedrock API
+            response = self.client.invoke_model(
+                body=body,
+                modelId=self.name,
+                accept="application/json",
+                contentType="application/json",
+            )
+
+            # format output based on provider
+            response_body = json.loads(response.get("body").read())
+            if provider == "cohere":
+                return response_body.get("embeddings")[0]
+            else:
+                # includes common provider == "amazon"
+                return response_body.get("embedding")
+        except Exception as e:
+            help_txt = """
+                boto3 client failed to invoke the bedrock API. In case of
+                AWS credentials error:
+                    - Please check your AWS credentials and ensure that you have access.
+                    You can set up aws credentials using `aws configure` command and
+                    verify by running `aws sts get-caller-identity` in your terminal.
+                """
+            raise ValueError(f"Error raised by boto3 client: {e}. \n {help_txt}")
+
+    @cached_property
+    def client(self):
+        """Create a boto3 client for Amazon Bedrock service
+
+        Returns
+        -------
+        boto3.client
+            The boto3 client for Amazon Bedrock service
+        """
+        botocore = attempt_import_or_raise("botocore")
+        boto3 = attempt_import_or_raise("boto3")
+
+        session_kwargs = {"region_name": self.region}
+        client_kwargs = {**session_kwargs}
+
+        if self.profile_name:
+            session_kwargs["profile_name"] = self.profile_name
+
+        retry_config = botocore.config.Config(
+            region_name=self.region,
+            retries={
+                "max_attempts": 0,  # disable this as retries retries are handled
+                "mode": "standard",
+            },
+        )
+        session = (
+            boto3.Session(**session_kwargs) if self.profile_name else boto3.Session()
+        )
+        if self.assumed_role:  # if not using default credentials
+            sts = session.client("sts")
+            response = sts.assume_role(
+                RoleArn=str(self.assumed_role),
+                RoleSessionName=self.role_session_name,
+            )
+            client_kwargs["aws_access_key_id"] = response["Credentials"]["AccessKeyId"]
+            client_kwargs["aws_secret_access_key"] = response["Credentials"][
+                "SecretAccessKey"
+            ]
+            client_kwargs["aws_session_token"] = response["Credentials"]["SessionToken"]
+
+        service_name = "bedrock-runtime"
+
+        bedrock_client = session.client(
+            service_name=service_name, config=retry_config, **client_kwargs
+        )
+
+        return bedrock_client

python/lancedb/embeddings/cohere.py

@@ -16,6 +16,7 @@ from typing import ClassVar, List, Union
 
 import numpy as np
 
+from ..util import attempt_import_or_raise
 from .base import TextEmbeddingFunction
 from .registry import register
 from .utils import api_key_not_found_help
@@ -84,7 +85,7 @@ class CohereEmbeddingFunction(TextEmbeddingFunction):
         return [emb for emb in rs.embeddings]
 
     def _init_client(self):
-        cohere = self.safe_import("cohere")
+        cohere = attempt_import_or_raise("cohere")
         if CohereEmbeddingFunction.client is None:
             if os.environ.get("COHERE_API_KEY") is None:
                 api_key_not_found_help("cohere")

python/lancedb/embeddings/gemini_text.py

@@ -19,6 +19,7 @@ import numpy as np
 
 from lancedb.pydantic import PYDANTIC_VERSION
 
+from ..util import attempt_import_or_raise
 from .base import TextEmbeddingFunction
 from .registry import register
 from .utils import TEXT, api_key_not_found_help
@@ -134,7 +135,7 @@ class GeminiText(TextEmbeddingFunction):
 
     @cached_property
     def client(self):
-        genai = self.safe_import("google.generativeai", "google.generativeai")
+        genai = attempt_import_or_raise("google.generativeai", "google.generativeai")
 
         if not os.environ.get("GOOGLE_API_KEY"):
             api_key_not_found_help("google")

python/lancedb/embeddings/gte.py

@@ -0,0 +1,131 @@
+#  Copyright (c) 2023. LanceDB Developers
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+from typing import List, Union
+
+import numpy as np
+
+from ..util import attempt_import_or_raise
+from .base import TextEmbeddingFunction
+from .registry import register
+from .utils import weak_lru
+
+
+@register("gte-text")
+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:
+        pip install mlx
+
+    Parameters
+    ----------
+    name: str, default "thenlper/gte-large"
+        The name of the model to use.
+    device: str, default "cpu"
+        Sets the device type for the model.
+    normalize: str, default "True"
+        Controls normalize param in encode function for the transformer.
+    mlx: bool, default False
+        Controls which model to use. False for gte-large,True for the mlx version.
+
+    Examples
+    --------
+    import lancedb
+    import lancedb.embeddings.gte
+    from lancedb.embeddings import get_registry
+    from lancedb.pydantic import LanceModel, Vector
+    import pandas as pd
+
+    model = get_registry().get("gte-text").create() # mlx=True for Apple silicon
+    class TextModel(LanceModel):
+        text: str = model.SourceField()
+        vector: Vector(model.ndims()) = model.VectorField()
+
+    df = pd.DataFrame({"text": ["hi hello sayonara", "goodbye world"]})
+    db = lancedb.connect("~/.lancedb")
+    tbl = db.create_table("test", schema=TextModel, mode="overwrite")
+
+    tbl.add(df)
+    rs = tbl.search("hello").limit(1).to_pandas()
+
+    """
+
+    name: str = "thenlper/gte-large"
+    device: str = "cpu"
+    normalize: bool = True
+    mlx: bool = False
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self._ndims = None
+        if kwargs:
+            self.mlx = kwargs.get("mlx", False)
+            if self.mlx is True:
+                self.name = "gte-mlx"
+
+    @property
+    def embedding_model(self):
+        """
+        Get the embedding model specified by the flag,
+        name and device. This is cached so that the model is only loaded
+        once per process.
+        """
+        return self.get_embedding_model()
+
+    def ndims(self):
+        if self.mlx is True:
+            self._ndims = self.embedding_model.dims
+        if self._ndims is None:
+            self._ndims = len(self.generate_embeddings("foo")[0])
+        return self._ndims
+
+    def generate_embeddings(
+        self, texts: Union[List[str], np.ndarray]
+    ) -> List[np.array]:
+        """
+        Get the embeddings for the given texts.
+
+        Parameters
+        ----------
+        texts: list[str] or np.ndarray (of str)
+            The texts to embed
+        """
+        if self.mlx is True:
+            return self.embedding_model.run(list(texts)).tolist()
+
+        return self.embedding_model.encode(
+            list(texts),
+            convert_to_numpy=True,
+            normalize_embeddings=self.normalize,
+        ).tolist()
+
+    @weak_lru(maxsize=1)
+    def get_embedding_model(self):
+        """
+        Get the embedding model specified by the flag,
+        name and device. This is cached so that the model is only loaded
+        once per process.
+        """
+        if self.mlx is True:
+            from .gte_mlx_model import Model
+
+            return Model()
+        else:
+            sentence_transformers = attempt_import_or_raise(
+                "sentence_transformers", "sentence-transformers"
+            )
+            return sentence_transformers.SentenceTransformer(
+                self.name, device=self.device
+            )

python/lancedb/embeddings/gte_mlx_model.py

@@ -0,0 +1,154 @@
+import json
+from typing import List, Optional
+
+import numpy as np
+from huggingface_hub import snapshot_download
+from pydantic import BaseModel
+from transformers import BertTokenizer
+
+try:
+    import mlx.core as mx
+    import mlx.nn as nn
+except ImportError:
+    raise ImportError("You need to install MLX to use this model use - pip install mlx")
+
+
+def average_pool(last_hidden_state: mx.array, attention_mask: mx.array) -> mx.array:
+    last_hidden = mx.multiply(last_hidden_state, attention_mask[..., None])
+    return last_hidden.sum(axis=1) / attention_mask.sum(axis=1)[..., None]
+
+
+class ModelConfig(BaseModel):
+    dim: int = 1024
+    num_attention_heads: int = 16
+    num_hidden_layers: int = 24
+    vocab_size: int = 30522
+    attention_probs_dropout_prob: float = 0.1
+    hidden_dropout_prob: float = 0.1
+    layer_norm_eps: float = 1e-12
+    max_position_embeddings: int = 512
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    A transformer encoder layer with (the original BERT) post-normalization.
+    """
+
+    def __init__(
+        self,
+        dims: int,
+        num_heads: int,
+        mlp_dims: Optional[int] = None,
+        layer_norm_eps: float = 1e-12,
+    ):
+        super().__init__()
+        mlp_dims = mlp_dims or dims * 4
+        self.attention = nn.MultiHeadAttention(dims, num_heads, bias=True)
+        self.ln1 = nn.LayerNorm(dims, eps=layer_norm_eps)
+        self.ln2 = nn.LayerNorm(dims, eps=layer_norm_eps)
+        self.linear1 = nn.Linear(dims, mlp_dims)
+        self.linear2 = nn.Linear(mlp_dims, dims)
+        self.gelu = nn.GELU()
+
+    def __call__(self, x, mask):
+        attention_out = self.attention(x, x, x, mask)
+        add_and_norm = self.ln1(x + attention_out)
+
+        ff = self.linear1(add_and_norm)
+        ff_gelu = self.gelu(ff)
+        ff_out = self.linear2(ff_gelu)
+        x = self.ln2(ff_out + add_and_norm)
+
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self, num_layers: int, dims: int, num_heads: int, mlp_dims: Optional[int] = None
+    ):
+        super().__init__()
+        self.layers = [
+            TransformerEncoderLayer(dims, num_heads, mlp_dims)
+            for i in range(num_layers)
+        ]
+
+    def __call__(self, x, mask):
+        for layer in self.layers:
+            x = layer(x, mask)
+
+        return x
+
+
+class BertEmbeddings(nn.Module):
+    def __init__(self, config: ModelConfig):
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.dim)
+        self.token_type_embeddings = nn.Embedding(2, config.dim)
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.dim
+        )
+        self.norm = nn.LayerNorm(config.dim, eps=config.layer_norm_eps)
+
+    def __call__(self, input_ids: mx.array, token_type_ids: mx.array) -> mx.array:
+        words = self.word_embeddings(input_ids)
+        position = self.position_embeddings(
+            mx.broadcast_to(mx.arange(input_ids.shape[1]), input_ids.shape)
+        )
+        token_types = self.token_type_embeddings(token_type_ids)
+
+        embeddings = position + words + token_types
+        return self.norm(embeddings)
+
+
+class Bert(nn.Module):
+    def __init__(self, config: ModelConfig):
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = TransformerEncoder(
+            num_layers=config.num_hidden_layers,
+            dims=config.dim,
+            num_heads=config.num_attention_heads,
+        )
+        self.pooler = nn.Linear(config.dim, config.dim)
+
+    def __call__(
+        self,
+        input_ids: mx.array,
+        token_type_ids: mx.array,
+        attention_mask: mx.array = None,
+    ) -> tuple[mx.array, mx.array]:
+        x = self.embeddings(input_ids, token_type_ids)
+
+        if attention_mask is not None:
+            # convert 0's to -infs, 1's to 0's, and make it broadcastable
+            attention_mask = mx.log(attention_mask)
+            attention_mask = mx.expand_dims(attention_mask, (1, 2))
+
+        y = self.encoder(x, attention_mask)
+        return y, mx.tanh(self.pooler(y[:, 0]))
+
+
+class Model:
+    def __init__(self) -> None:
+        # get converted embedding model
+        model_path = snapshot_download(repo_id="vegaluisjose/mlx-rag")
+        with open(f"{model_path}/config.json") as f:
+            model_config = ModelConfig(**json.load(f))
+        self.dims = model_config.dim
+        self.model = Bert(model_config)
+        self.model.load_weights(f"{model_path}/model.npz")
+        self.tokenizer = BertTokenizer.from_pretrained("thenlper/gte-large")
+        self.embeddings = []
+
+    def run(self, input_text: List[str]) -> mx.array:
+        tokens = self.tokenizer(input_text, return_tensors="np", padding=True)
+        tokens = {key: mx.array(v) for key, v in tokens.items()}
+
+        last_hidden_state, _ = self.model(**tokens)
+
+        embeddings = average_pool(
+            last_hidden_state, tokens["attention_mask"].astype(mx.float32)
+        )
+        self.embeddings = (
+            embeddings / mx.linalg.norm(embeddings, ord=2, axis=1)[..., None]
+        )
+
+        return np.array(embeddings.astype(mx.float32))

python/lancedb/embeddings/instructor.py

@@ -14,6 +14,7 @@ from typing import List
 
 import numpy as np
 
+from ..util import attempt_import_or_raise
 from .base import TextEmbeddingFunction
 from .registry import register
 from .utils import TEXT, weak_lru
@@ -102,9 +103,9 @@ class InstructorEmbeddingFunction(TextEmbeddingFunction):
     # convert_to_numpy: bool = True # Hardcoding this as numpy can be ingested directly
 
     source_instruction: str = "represent the document for retrieval"
-    query_instruction: str = (
-        "represent the document for retrieving the most similar documents"
-    )
+    query_instruction: (
+        str
+    ) = "represent the document for retrieving the most similar documents"
 
     @weak_lru(maxsize=1)
     def ndims(self):
@@ -131,10 +132,10 @@ class InstructorEmbeddingFunction(TextEmbeddingFunction):
 
     @weak_lru(maxsize=1)
     def get_model(self):
-        instructor_embedding = self.safe_import(
+        instructor_embedding = attempt_import_or_raise(
             "InstructorEmbedding", "InstructorEmbedding"
         )
-        torch = self.safe_import("torch", "torch")
+        torch = attempt_import_or_raise("torch", "torch")
 
         model = instructor_embedding.INSTRUCTOR(self.name)
         if self.quantize:

python/lancedb/embeddings/open_clip.py

@@ -21,6 +21,7 @@ import pyarrow as pa
 from pydantic import PrivateAttr
 from tqdm import tqdm
 
+from ..util import attempt_import_or_raise
 from .base import EmbeddingFunction
 from .registry import register
 from .utils import IMAGES, url_retrieve
@@ -50,7 +51,7 @@ class OpenClipEmbeddings(EmbeddingFunction):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        open_clip = self.safe_import("open_clip", "open-clip")
+        open_clip = attempt_import_or_raise("open_clip", "open-clip")
         model, _, preprocess = open_clip.create_model_and_transforms(
             self.name, pretrained=self.pretrained
         )
@@ -78,14 +79,14 @@ class OpenClipEmbeddings(EmbeddingFunction):
         if isinstance(query, str):
             return [self.generate_text_embeddings(query)]
         else:
-            PIL = self.safe_import("PIL", "pillow")
+            PIL = attempt_import_or_raise("PIL", "pillow")
             if isinstance(query, PIL.Image.Image):
                 return [self.generate_image_embedding(query)]
             else:
                 raise TypeError("OpenClip supports str or PIL Image as query")
 
     def generate_text_embeddings(self, text: str) -> np.ndarray:
-        torch = self.safe_import("torch")
+        torch = attempt_import_or_raise("torch")
         text = self.sanitize_input(text)
         text = self._tokenizer(text)
         text.to(self.device)
@@ -144,7 +145,7 @@ class OpenClipEmbeddings(EmbeddingFunction):
             The image to embed. If the image is a str, it is treated as a uri.
             If the image is bytes, it is treated as the raw image bytes.
         """
-        torch = self.safe_import("torch")
+        torch = attempt_import_or_raise("torch")
         # TODO handle retry and errors for https
         image = self._to_pil(image)
         image = self._preprocess(image).unsqueeze(0)
@@ -152,7 +153,7 @@ class OpenClipEmbeddings(EmbeddingFunction):
             return self._encode_and_normalize_image(image)
 
     def _to_pil(self, image: Union[str, bytes]):
-        PIL = self.safe_import("PIL", "pillow")
+        PIL = attempt_import_or_raise("PIL", "pillow")
         if isinstance(image, bytes):
             return PIL.Image.open(io.BytesIO(image))
         if isinstance(image, PIL.Image.Image):

python/lancedb/embeddings/openai.py

@@ -12,10 +12,11 @@
 #  limitations under the License.
 import os
 from functools import cached_property
-from typing import List, Union
+from typing import List, Optional, Union
 
 import numpy as np
 
+from ..util import attempt_import_or_raise
 from .base import TextEmbeddingFunction
 from .registry import register
 from .utils import api_key_not_found_help
@@ -30,10 +31,21 @@ class OpenAIEmbeddings(TextEmbeddingFunction):
     """
 
     name: str = "text-embedding-ada-002"
+    dim: Optional[int] = None
 
     def ndims(self):
-        # TODO don't hardcode this
-        return 1536
+        return self._ndims
+
+    @cached_property
+    def _ndims(self):
+        if self.name == "text-embedding-ada-002":
+            return 1536
+        elif self.name == "text-embedding-3-large":
+            return self.dim or 3072
+        elif self.name == "text-embedding-3-small":
+            return self.dim or 1536
+        else:
+            raise ValueError(f"Unknown model name {self.name}")
 
     def generate_embeddings(
         self, texts: Union[List[str], np.ndarray]
@@ -47,12 +59,17 @@ class OpenAIEmbeddings(TextEmbeddingFunction):
             The texts to embed
         """
         # TODO retry, rate limit, token limit
-        rs = self._openai_client.embeddings.create(input=texts, model=self.name)
+        if self.name == "text-embedding-ada-002":
+            rs = self._openai_client.embeddings.create(input=texts, model=self.name)
+        else:
+            rs = self._openai_client.embeddings.create(
+                input=texts, model=self.name, dimensions=self.ndims()
+            )
         return [v.embedding for v in rs.data]
 
     @cached_property
     def _openai_client(self):
-        openai = self.safe_import("openai")
+        openai = attempt_import_or_raise("openai")
 
         if not os.environ.get("OPENAI_API_KEY"):
             api_key_not_found_help("openai")

python/lancedb/embeddings/sentence_transformers.py

@@ -14,6 +14,7 @@ from typing import List, Union
 
 import numpy as np
 
+from ..util import attempt_import_or_raise
 from .base import TextEmbeddingFunction
 from .registry import register
 from .utils import weak_lru
@@ -75,7 +76,7 @@ class SentenceTransformerEmbeddings(TextEmbeddingFunction):
 
         TODO: use lru_cache instead with a reasonable/configurable maxsize
         """
-        sentence_transformers = self.safe_import(
+        sentence_transformers = attempt_import_or_raise(
             "sentence_transformers", "sentence-transformers"
         )
         return sentence_transformers.SentenceTransformer(self.name, device=self.device)

python/lancedb/merge.py

@@ -0,0 +1,107 @@
+#  Copyright 2023 LanceDB Developers
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, List, Optional
+
+if TYPE_CHECKING:
+    from .common import DATA
+
+
+class LanceMergeInsertBuilder(object):
+    """Builder for a LanceDB merge insert operation
+
+    See [`merge_insert`][lancedb.table.Table.merge_insert] for
+    more context
+    """
+
+    def __init__(self, table: "Table", on: List[str]):  # noqa: F821
+        # Do not put a docstring here.  This method should be hidden
+        # from API docs.  Users should use merge_insert to create
+        # this object.
+        self._table = table
+        self._on = on
+        self._when_matched_update_all = False
+        self._when_matched_update_all_condition = None
+        self._when_not_matched_insert_all = False
+        self._when_not_matched_by_source_delete = False
+        self._when_not_matched_by_source_condition = None
+
+    def when_matched_update_all(
+        self, *, where: Optional[str] = None
+    ) -> LanceMergeInsertBuilder:
+        """
+        Rows that exist in both the source table (new data) and
+        the target table (old data) will be updated, replacing
+        the old row with the corresponding matching row.
+
+        If there are multiple matches then the behavior is undefined.
+        Currently this causes multiple copies of the row to be created
+        but that behavior is subject to change.
+        """
+        self._when_matched_update_all = True
+        self._when_matched_update_all_condition = where
+        return self
+
+    def when_not_matched_insert_all(self) -> LanceMergeInsertBuilder:
+        """
+        Rows that exist only in the source table (new data) should
+        be inserted into the target table.
+        """
+        self._when_not_matched_insert_all = True
+        return self
+
+    def when_not_matched_by_source_delete(
+        self, condition: Optional[str] = None
+    ) -> LanceMergeInsertBuilder:
+        """
+        Rows that exist only in the target table (old data) will be
+        deleted.  An optional condition can be provided to limit what
+        data is deleted.
+
+        Parameters
+        ----------
+        condition: Optional[str], default None
+            If None then all such rows will be deleted.  Otherwise the
+            condition will be used as an SQL filter to limit what rows
+            are deleted.
+        """
+        self._when_not_matched_by_source_delete = True
+        if condition is not None:
+            self._when_not_matched_by_source_condition = condition
+        return self
+
+    def execute(
+        self,
+        new_data: DATA,
+        on_bad_vectors: str = "error",
+        fill_value: float = 0.0,
+    ):
+        """
+        Executes the merge insert operation
+
+        Nothing is returned but the [`Table`][lancedb.table.Table] is updated
+
+        Parameters
+        ----------
+        new_data: DATA
+            New records which will be matched against the existing records
+            to potentially insert or update into the table.  This parameter
+            can be anything you use for [`add`][lancedb.table.Table.add]
+        on_bad_vectors: str, default "error"
+            What to do if any of the vectors are not the same size or contains NaNs.
+            One of "error", "drop", "fill".
+        fill_value: float, default 0.
+            The value to use when filling vectors. Only used if on_bad_vectors="fill".
+        """
+        self._table._do_merge(self, new_data, on_bad_vectors, fill_value)

python/lancedb/pydantic.py

@@ -304,7 +304,7 @@ class LanceModel(pydantic.BaseModel):
     ...     name: str
     ...     vector: Vector(2)
     ...
-    >>> db = lancedb.connect("/tmp")
+    >>> db = lancedb.connect("./example")
     >>> table = db.create_table("test", schema=TestModel.to_arrow_schema())
     >>> table.add([
     ...     TestModel(name="test", vector=[1.0, 2.0])

python/lancedb/query.py

@@ -14,8 +14,9 @@
 from __future__ import annotations
 
 from abc import ABC, abstractmethod
+from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
-from typing import TYPE_CHECKING, List, Literal, Optional, Type, Union
+from typing import TYPE_CHECKING, List, Literal, Optional, Tuple, Type, Union
 
 import deprecation
 import numpy as np
@@ -23,7 +24,9 @@ import pyarrow as pa
 import pydantic
 
 from . import __version__
-from .common import VECTOR_COLUMN_NAME
+from .common import VEC
+from .rerankers.base import Reranker
+from .rerankers.linear_combination import LinearCombinationReranker
 from .util import safe_import_pandas
 
 if TYPE_CHECKING:
@@ -72,7 +75,7 @@ class Query(pydantic.BaseModel):
           tuning advice.
     """
 
-    vector_column: str = VECTOR_COLUMN_NAME
+    vector_column: Optional[str] = None
 
     # vector to search for
     vector: Union[List[float], List[List[float]]]
@@ -99,6 +102,8 @@ class Query(pydantic.BaseModel):
     # Refine factor.
     refine_factor: Optional[int] = None
 
+    with_row_id: bool = False
+
 
 class LanceQueryBuilder(ABC):
     """Build LanceDB query based on specific query type:
@@ -109,19 +114,26 @@ class LanceQueryBuilder(ABC):
     def create(
         cls,
         table: "Table",
-        query: Optional[Union[np.ndarray, str, "PIL.Image.Image"]],
+        query: Optional[Union[np.ndarray, str, "PIL.Image.Image", Tuple]],
         query_type: str,
         vector_column_name: str,
     ) -> LanceQueryBuilder:
         if query is None:
             return LanceEmptyQueryBuilder(table)
 
-        # convert "auto" query_type to "vector" or "fts"
-        # and convert the query to vector if needed
+        if query_type == "hybrid":
+            # hybrid fts and vector query
+            return LanceHybridQueryBuilder(table, query, vector_column_name)
+
+        # convert "auto" query_type to "vector", "fts"
+        # or "hybrid" and convert the query to vector if needed
         query, query_type = cls._resolve_query(
             table, query, query_type, vector_column_name
         )
 
+        if query_type == "hybrid":
+            return LanceHybridQueryBuilder(table, query, vector_column_name)
+
         if isinstance(query, str):
             # fts
             return LanceFtsQueryBuilder(table, query)
@@ -144,17 +156,13 @@ class LanceQueryBuilder(ABC):
                 raise TypeError(f"'fts' queries must be a string: {type(query)}")
             return query, query_type
         elif query_type == "vector":
-            if not isinstance(query, (list, np.ndarray)):
-                conf = table.embedding_functions.get(vector_column_name)
-                if conf is not None:
-                    query = conf.function.compute_query_embeddings_with_retry(query)[0]
-                else:
-                    msg = f"No embedding function for {vector_column_name}"
-                    raise ValueError(msg)
+            query = cls._query_to_vector(table, query, vector_column_name)
             return query, query_type
         elif query_type == "auto":
             if isinstance(query, (list, np.ndarray)):
                 return query, "vector"
+            if isinstance(query, tuple):
+                return query, "hybrid"
             else:
                 conf = table.embedding_functions.get(vector_column_name)
                 if conf is not None:
@@ -167,11 +175,23 @@ class LanceQueryBuilder(ABC):
                 f"Invalid query_type, must be 'vector', 'fts', or 'auto': {query_type}"
             )
 
+    @classmethod
+    def _query_to_vector(cls, table, query, vector_column_name):
+        if isinstance(query, (list, np.ndarray)):
+            return query
+        conf = table.embedding_functions.get(vector_column_name)
+        if conf is not None:
+            return conf.function.compute_query_embeddings_with_retry(query)[0]
+        else:
+            msg = f"No embedding function for {vector_column_name}"
+            raise ValueError(msg)
+
     def __init__(self, table: "Table"):
         self._table = table
         self._limit = 10
         self._columns = None
         self._where = None
+        self._with_row_id = False
 
     @deprecation.deprecated(
         deprecated_in="0.3.1",
@@ -341,6 +361,22 @@ class LanceQueryBuilder(ABC):
         self._prefilter = prefilter
         return self
 
+    def with_row_id(self, with_row_id: bool) -> LanceQueryBuilder:
+        """Set whether to return row ids.
+
+        Parameters
+        ----------
+        with_row_id: bool
+            If True, return _rowid column in the results.
+
+        Returns
+        -------
+        LanceQueryBuilder
+            The LanceQueryBuilder object.
+        """
+        self._with_row_id = with_row_id
+        return self
+
 
 class LanceVectorQueryBuilder(LanceQueryBuilder):
     """
@@ -367,7 +403,7 @@ class LanceVectorQueryBuilder(LanceQueryBuilder):
         self,
         table: "Table",
         query: Union[np.ndarray, list, "PIL.Image.Image"],
-        vector_column: str = VECTOR_COLUMN_NAME,
+        vector_column: str,
     ):
         super().__init__(table)
         self._query = query
@@ -459,6 +495,7 @@ class LanceVectorQueryBuilder(LanceQueryBuilder):
             nprobes=self._nprobes,
             refine_factor=self._refine_factor,
             vector_column=self._vector_column,
+            with_row_id=self._with_row_id,
         )
         return self._table._execute_query(query)
 
@@ -568,6 +605,10 @@ class LanceFtsQueryBuilder(LanceQueryBuilder):
                     ds = lance.write_dataset(output_tbl, tmp)
                     output_tbl = ds.to_table(filter=self._where)
 
+        if self._with_row_id:
+            # Need to set this to uint explicitly as vector results are in uint64
+            row_ids = pa.array(row_ids, type=pa.uint64())
+            output_tbl = output_tbl.append_column("_rowid", row_ids)
         return output_tbl
 
 
@@ -579,3 +620,265 @@ class LanceEmptyQueryBuilder(LanceQueryBuilder):
             filter=self._where,
             limit=self._limit,
         )
+
+
+class LanceHybridQueryBuilder(LanceQueryBuilder):
+    def __init__(self, table: "Table", query: str, vector_column: str):
+        super().__init__(table)
+        self._validate_fts_index()
+        vector_query, fts_query = self._validate_query(query)
+        self._fts_query = LanceFtsQueryBuilder(table, fts_query)
+        vector_query = self._query_to_vector(table, vector_query, vector_column)
+        self._vector_query = LanceVectorQueryBuilder(table, vector_query, vector_column)
+        self._norm = "score"
+        self._reranker = LinearCombinationReranker(weight=0.7, fill=1.0)
+
+    def _validate_fts_index(self):
+        if self._table._get_fts_index_path() is None:
+            raise ValueError(
+                "Please create a full-text search index " "to perform hybrid search."
+            )
+
+    def _validate_query(self, query):
+        # Temp hack to support vectorized queries for hybrid search
+        if isinstance(query, str):
+            return query, query
+        elif isinstance(query, tuple):
+            if len(query) != 2:
+                raise ValueError(
+                    "The query must be a tuple of (vector_query, fts_query)."
+                )
+            if not isinstance(query[0], (list, np.ndarray, pa.Array, pa.ChunkedArray)):
+                raise ValueError(f"The vector query must be one of {VEC}.")
+            if not isinstance(query[1], str):
+                raise ValueError("The fts query must be a string.")
+            return query[0], query[1]
+        else:
+            raise ValueError(
+                "The query must be either a string or a tuple of (vector, string)."
+            )
+
+    def to_arrow(self) -> pa.Table:
+        with ThreadPoolExecutor() as executor:
+            fts_future = executor.submit(self._fts_query.with_row_id(True).to_arrow)
+            vector_future = executor.submit(
+                self._vector_query.with_row_id(True).to_arrow
+            )
+            fts_results = fts_future.result()
+            vector_results = vector_future.result()
+
+        # convert to ranks first if needed
+        if self._norm == "rank":
+            vector_results = self._rank(vector_results, "_distance")
+            fts_results = self._rank(fts_results, "score")
+        # normalize the scores to be between 0 and 1, 0 being most relevant
+        vector_results = self._normalize_scores(vector_results, "_distance")
+
+        # In fts higher scores represent relevance. Not inverting them here as
+        # rerankers might need to preserve this score to support `return_score="all"`
+        fts_results = self._normalize_scores(fts_results, "score")
+
+        results = self._reranker.rerank_hybrid(
+            self._fts_query._query, vector_results, fts_results
+        )
+
+        if not isinstance(results, pa.Table):  # Enforce type
+            raise TypeError(
+                f"rerank_hybrid must return a pyarrow.Table, got {type(results)}"
+            )
+
+        # apply limit after reranking
+        results = results.slice(length=self._limit)
+
+        if not self._with_row_id:
+            results = results.drop(["_rowid"])
+        return results
+
+    def _rank(self, results: pa.Table, column: str, ascending: bool = True):
+        if len(results) == 0:
+            return results
+        # Get the _score column from results
+        scores = results.column(column).to_numpy()
+        sort_indices = np.argsort(scores)
+        if not ascending:
+            sort_indices = sort_indices[::-1]
+        ranks = np.empty_like(sort_indices)
+        ranks[sort_indices] = np.arange(len(scores)) + 1
+        # replace the _score column with the ranks
+        _score_idx = results.column_names.index(column)
+        results = results.set_column(
+            _score_idx, column, pa.array(ranks, type=pa.float32())
+        )
+        return results
+
+    def _normalize_scores(self, results: pa.Table, column: str, invert=False):
+        if len(results) == 0:
+            return results
+        # Get the _score column from results
+        scores = results.column(column).to_numpy()
+        # normalize the scores by subtracting the min and dividing by the max
+        max, min = np.max(scores), np.min(scores)
+        if np.isclose(max, min):
+            rng = max
+        else:
+            rng = max - min
+        scores = (scores - min) / rng
+        if invert:
+            scores = 1 - scores
+        # replace the _score column with the ranks
+        _score_idx = results.column_names.index(column)
+        results = results.set_column(
+            _score_idx, column, pa.array(scores, type=pa.float32())
+        )
+        return results
+
+    def rerank(
+        self,
+        normalize="score",
+        reranker: Reranker = LinearCombinationReranker(weight=0.7, fill=1.0),
+    ) -> LanceHybridQueryBuilder:
+        """
+        Rerank the hybrid search results using the specified reranker. The reranker
+        must be an instance of Reranker class.
+
+        Parameters
+        ----------
+        normalize: str, default "score"
+            The method to normalize the scores. Can be "rank" or "score". If "rank",
+            the scores are converted to ranks and then normalized. If "score", the
+            scores are normalized directly.
+        reranker: Reranker, default LinearCombinationReranker(weight=0.7, fill=1.0)
+            The reranker to use. Must be an instance of Reranker class.
+        Returns
+        -------
+        LanceHybridQueryBuilder
+            The LanceHybridQueryBuilder object.
+        """
+        if normalize not in ["rank", "score"]:
+            raise ValueError("normalize must be 'rank' or 'score'.")
+        if reranker and not isinstance(reranker, Reranker):
+            raise ValueError("reranker must be an instance of Reranker class.")
+
+        self._norm = normalize
+        self._reranker = reranker
+
+        return self
+
+    def limit(self, limit: int) -> LanceHybridQueryBuilder:
+        """
+        Set the maximum number of results to return for both vector and fts search
+        components.
+
+        Parameters
+        ----------
+        limit: int
+            The maximum number of results to return.
+
+        Returns
+        -------
+        LanceHybridQueryBuilder
+            The LanceHybridQueryBuilder object.
+        """
+        self._vector_query.limit(limit)
+        self._fts_query.limit(limit)
+        self._limit = limit
+
+        return self
+
+    def select(self, columns: list) -> LanceHybridQueryBuilder:
+        """
+        Set the columns to return for both vector and fts search.
+
+        Parameters
+        ----------
+        columns: list
+            The columns to return.
+
+        Returns
+        -------
+        LanceHybridQueryBuilder
+            The LanceHybridQueryBuilder object.
+        """
+        self._vector_query.select(columns)
+        self._fts_query.select(columns)
+        return self
+
+    def where(self, where: str, prefilter: bool = False) -> LanceHybridQueryBuilder:
+        """
+        Set the where clause for both vector and fts search.
+
+        Parameters
+        ----------
+        where: str
+            The where clause which is a valid SQL where clause. See
+            `Lance filter pushdown <https://lancedb.github.io/lance/read_and_write.html#filter-push-down>`_
+            for valid SQL expressions.
+
+        prefilter: bool, default False
+            If True, apply the filter before vector search, otherwise the
+            filter is applied on the result of vector search.
+
+        Returns
+        -------
+        LanceHybridQueryBuilder
+            The LanceHybridQueryBuilder object.
+        """
+
+        self._vector_query.where(where, prefilter=prefilter)
+        self._fts_query.where(where)
+        return self
+
+    def metric(self, metric: Literal["L2", "cosine"]) -> LanceHybridQueryBuilder:
+        """
+        Set the distance metric to use for vector search.
+
+        Parameters
+        ----------
+        metric: "L2" or "cosine"
+            The distance metric to use. By default "L2" is used.
+
+        Returns
+        -------
+        LanceHybridQueryBuilder
+            The LanceHybridQueryBuilder object.
+        """
+        self._vector_query.metric(metric)
+        return self
+
+    def nprobes(self, nprobes: int) -> LanceHybridQueryBuilder:
+        """
+        Set the number of probes to use for vector search.
+
+        Higher values will yield better recall (more likely to find vectors if
+        they exist) at the expense of latency.
+
+        Parameters
+        ----------
+        nprobes: int
+            The number of probes to use.
+
+        Returns
+        -------
+        LanceHybridQueryBuilder
+            The LanceHybridQueryBuilder object.
+        """
+        self._vector_query.nprobes(nprobes)
+        return self
+
+    def refine_factor(self, refine_factor: int) -> LanceHybridQueryBuilder:
+        """
+        Refine the vector search results by reading extra elements and
+        re-ranking them in memory.
+
+        Parameters
+        ----------
+        refine_factor: int
+            The refine factor to use.
+
+        Returns
+        -------
+        LanceHybridQueryBuilder
+            The LanceHybridQueryBuilder object.
+        """
+        self._vector_query.refine_factor(refine_factor)
+        return self

python/lancedb/remote/client.py

@@ -13,6 +13,8 @@
 
 
 import functools
+import logging
+import os
 from typing import Any, Callable, Dict, List, Optional, Union
 from urllib.parse import urljoin
 
@@ -20,6 +22,8 @@ import attrs
 import pyarrow as pa
 import requests
 from pydantic import BaseModel
+from requests.adapters import HTTPAdapter
+from urllib3 import Retry
 
 from lancedb.common import Credential
 from lancedb.remote import VectorQuery, VectorQueryResult
@@ -57,6 +61,10 @@ class RestfulLanceDBClient:
     @functools.cached_property
     def session(self) -> requests.Session:
         sess = requests.Session()
+
+        retry_adapter_instance = retry_adapter(retry_adapter_options())
+        sess.mount(urljoin(self.url, "/v1/table/"), retry_adapter_instance)
+
         adapter_class = LanceDBClientHTTPAdapterFactory()
         sess.mount("https://", adapter_class())
         return sess
@@ -109,7 +117,7 @@ class RestfulLanceDBClient:
             urljoin(self.url, uri),
             params=params,
             headers=self.headers,
-            timeout=(10.0, 300.0),
+            timeout=(120.0, 300.0),
         ) as resp:
             self._check_status(resp)
             return resp.json()
@@ -151,7 +159,7 @@ class RestfulLanceDBClient:
             urljoin(self.url, uri),
             headers=headers,
             params=params,
-            timeout=(10.0, 300.0),
+            timeout=(120.0, 300.0),
             **req_kwargs,
         ) as resp:
             self._check_status(resp)
@@ -170,3 +178,72 @@ class RestfulLanceDBClient:
         """Query a table."""
         tbl = self.post(f"/v1/table/{table_name}/query/", query, deserialize=_read_ipc)
         return VectorQueryResult(tbl)
+
+    def mount_retry_adapter_for_table(self, table_name: str) -> None:
+        """
+        Adds an http adapter to session that will retry retryable requests to the table.
+        """
+        retry_options = retry_adapter_options(methods=["GET", "POST"])
+        retry_adapter_instance = retry_adapter(retry_options)
+        session = self.session
+
+        session.mount(
+            urljoin(self.url, f"/v1/table/{table_name}/query/"), retry_adapter_instance
+        )
+        session.mount(
+            urljoin(self.url, f"/v1/table/{table_name}/describe/"),
+            retry_adapter_instance,
+        )
+        session.mount(
+            urljoin(self.url, f"/v1/table/{table_name}/index/list/"),
+            retry_adapter_instance,
+        )
+
+
+def retry_adapter_options(methods=["GET"]) -> Dict[str, Any]:
+    return {
+        "retries": int(os.environ.get("LANCE_CLIENT_MAX_RETRIES", "3")),
+        "connect_retries": int(os.environ.get("LANCE_CLIENT_CONNECT_RETRIES", "3")),
+        "read_retries": int(os.environ.get("LANCE_CLIENT_READ_RETRIES", "3")),
+        "backoff_factor": float(
+            os.environ.get("LANCE_CLIENT_RETRY_BACKOFF_FACTOR", "0.25")
+        ),
+        "backoff_jitter": float(
+            os.environ.get("LANCE_CLIENT_RETRY_BACKOFF_JITTER", "0.25")
+        ),
+        "statuses": [
+            int(i.strip())
+            for i in os.environ.get(
+                "LANCE_CLIENT_RETRY_STATUSES", "429, 500, 502, 503"
+            ).split(",")
+        ],
+        "methods": methods,
+    }
+
+
+def retry_adapter(options: Dict[str, Any]) -> HTTPAdapter:
+    total_retries = options["retries"]
+    connect_retries = options["connect_retries"]
+    read_retries = options["read_retries"]
+    backoff_factor = options["backoff_factor"]
+    backoff_jitter = options["backoff_jitter"]
+    statuses = options["statuses"]
+    methods = frozenset(options["methods"])
+    logging.debug(
+        f"Setting up retry adapter with {total_retries} retries,"  # noqa G003
+        + f"connect retries {connect_retries}, read retries {read_retries},"
+        + f"backoff factor {backoff_factor}, statuses {statuses}, "
+        + f"methods {methods}"
+    )
+
+    return HTTPAdapter(
+        max_retries=Retry(
+            total=total_retries,
+            connect=connect_retries,
+            read=read_retries,
+            backoff_factor=backoff_factor,
+            backoff_jitter=backoff_jitter,
+            status_forcelist=statuses,
+            allowed_methods=methods,
+        )
+    )

python/lancedb/remote/db.py

@@ -95,6 +95,8 @@ class RemoteDBConnection(DBConnection):
         """
         from .table import RemoteTable
 
+        self._client.mount_retry_adapter_for_table(name)
+
         # check if table exists
         try:
             self._client.post(f"/v1/table/{name}/describe/")
@@ -116,6 +118,7 @@ class RemoteDBConnection(DBConnection):
         schema: Optional[Union[pa.Schema, LanceModel]] = None,
         on_bad_vectors: str = "error",
         fill_value: float = 0.0,
+        mode: Optional[str] = None,
         embedding_functions: Optional[List[EmbeddingFunctionConfig]] = None,
     ) -> Table:
         """Create a [Table][lancedb.table.Table] in the database.
@@ -213,11 +216,13 @@ class RemoteDBConnection(DBConnection):
         if data is None and schema is None:
             raise ValueError("Either data or schema must be provided.")
         if embedding_functions is not None:
-            raise NotImplementedError(
-                "embedding_functions is not supported for remote databases."
+            logging.warning(
+                "embedding_functions is not yet supported on LanceDB Cloud."
                 "Please vote https://github.com/lancedb/lancedb/issues/626 "
                 "for this feature."
             )
+        if mode is not None:
+            logging.warning("mode is not yet supported on LanceDB Cloud.")
 
         if inspect.isclass(schema) and issubclass(schema, LanceModel):
             # convert LanceModel to pyarrow schema

python/lancedb/remote/table.py

@@ -11,6 +11,7 @@
 #  See the License for the specific language governing permissions and
 #  limitations under the License.
 
+import logging
 import uuid
 from functools import cached_property
 from typing import Dict, Optional, Union
@@ -19,10 +20,11 @@ import pyarrow as pa
 from lance import json_to_schema
 
 from lancedb.common import DATA, VEC, VECTOR_COLUMN_NAME
+from lancedb.merge import LanceMergeInsertBuilder
 
 from ..query import LanceVectorQueryBuilder
 from ..table import Query, Table, _sanitize_data
-from ..util import value_to_sql
+from ..util import inf_vector_column_query, value_to_sql
 from .arrow import to_ipc_binary
 from .client import ARROW_STREAM_CONTENT_TYPE
 from .db import RemoteDBConnection
@@ -36,6 +38,9 @@ class RemoteTable(Table):
     def __repr__(self) -> str:
         return f"RemoteTable({self._conn.db_name}.{self._name})"
 
+    def __len__(self) -> int:
+        self.count_rows(None)
+
     @cached_property
     def schema(self) -> pa.Schema:
         """The [Arrow Schema](https://arrow.apache.org/docs/python/api/datatypes.html#)
@@ -53,17 +58,17 @@ class RemoteTable(Table):
         return resp["version"]
 
     def to_arrow(self) -> pa.Table:
-        """to_arrow() is not supported on the LanceDB cloud"""
-        raise NotImplementedError("to_arrow() is not supported on the LanceDB cloud")
+        """to_arrow() is not yet supported on LanceDB cloud."""
+        raise NotImplementedError("to_arrow() is not yet supported on LanceDB cloud.")
 
     def to_pandas(self):
-        """to_pandas() is not supported on the LanceDB cloud"""
-        return NotImplementedError("to_pandas() is not supported on the LanceDB cloud")
+        """to_pandas() is not yet supported on LanceDB cloud."""
+        return NotImplementedError("to_pandas() is not yet supported on LanceDB cloud.")
 
     def create_scalar_index(self, *args, **kwargs):
         """Creates a scalar index"""
         return NotImplementedError(
-            "create_scalar_index() is not supported on the LanceDB cloud"
+            "create_scalar_index() is not yet supported on LanceDB cloud."
         )
 
     def create_index(
@@ -71,6 +76,10 @@ class RemoteTable(Table):
         metric="L2",
         vector_column_name: str = VECTOR_COLUMN_NAME,
         index_cache_size: Optional[int] = None,
+        num_partitions: Optional[int] = None,
+        num_sub_vectors: Optional[int] = None,
+        replace: Optional[bool] = None,
+        accelerator: Optional[str] = None,
     ):
         """Create an index on the table.
         Currently, the only parameters that matter are
@@ -104,6 +113,28 @@ class RemoteTable(Table):
         ... )
         >>> table.create_index("L2", "vector") # doctest: +SKIP
         """
+
+        if num_partitions is not None:
+            logging.warning(
+                "num_partitions is not supported on LanceDB cloud."
+                "This parameter will be tuned automatically."
+            )
+        if num_sub_vectors is not None:
+            logging.warning(
+                "num_sub_vectors is not supported on LanceDB cloud."
+                "This parameter will be tuned automatically."
+            )
+        if accelerator is not None:
+            logging.warning(
+                "GPU accelerator is not yet supported on LanceDB cloud."
+                "If you have 100M+ vectors to index,"
+                "please contact us at contact@lancedb.com"
+            )
+        if replace is not None:
+            logging.warning(
+                "replace is not supported on LanceDB cloud."
+                "Existing indexes will always be replaced."
+            )
         index_type = "vector"
 
         data = {
@@ -167,7 +198,9 @@ class RemoteTable(Table):
         )
 
     def search(
-        self, query: Union[VEC, str], vector_column_name: str = VECTOR_COLUMN_NAME
+        self,
+        query: Union[VEC, str],
+        vector_column_name: Optional[str] = None,
     ) -> LanceVectorQueryBuilder:
         """Create a search query to find the nearest neighbors
         of the given query vector. We currently support [vector search][search]
@@ -186,7 +219,7 @@ class RemoteTable(Table):
         ... ]
         >>> table = db.create_table("my_table", data) # doctest: +SKIP
         >>> query = [0.4, 1.4, 2.4]
-        >>> (table.search(query, vector_column_name="vector") # doctest: +SKIP
+        >>> (table.search(query) # doctest: +SKIP
         ...     .where("original_width > 1000", prefilter=True) # doctest: +SKIP
         ...     .select(["caption", "original_width"]) # doctest: +SKIP
         ...     .limit(2) # doctest: +SKIP
@@ -205,9 +238,14 @@ class RemoteTable(Table):
 
             - If None then the select/where/limit clauses are applied to filter
             the table
-        vector_column_name: str
+        vector_column_name: str, optional
             The name of the vector column to search.
-            *default "vector"*
+
+            - If not specified then the vector column is inferred from
+            the table schema
+
+            - If the table has multiple vector columns then the *vector_column_name*
+            needs to be specified. Otherwise, an error is raised.
 
         Returns
         -------
@@ -222,6 +260,8 @@ class RemoteTable(Table):
             - and also the "_distance" column which is the distance between the query
             vector and the returned vector.
         """
+        if vector_column_name is None:
+            vector_column_name = inf_vector_column_query(self.schema)
         return LanceVectorQueryBuilder(self, query, vector_column_name)
 
     def _execute_query(self, query: Query) -> pa.Table:
@@ -244,6 +284,51 @@ class RemoteTable(Table):
             result = self._conn._client.query(self._name, query)
             return result.to_arrow()
 
+    def _do_merge(
+        self,
+        merge: LanceMergeInsertBuilder,
+        new_data: DATA,
+        on_bad_vectors: str,
+        fill_value: float,
+    ):
+        data = _sanitize_data(
+            new_data,
+            self.schema,
+            metadata=None,
+            on_bad_vectors=on_bad_vectors,
+            fill_value=fill_value,
+        )
+        payload = to_ipc_binary(data)
+
+        params = {}
+        if len(merge._on) != 1:
+            raise ValueError(
+                "RemoteTable only supports a single on key in merge_insert"
+            )
+        params["on"] = merge._on[0]
+        params["when_matched_update_all"] = str(merge._when_matched_update_all).lower()
+        if merge._when_matched_update_all_condition is not None:
+            params[
+                "when_matched_update_all_filt"
+            ] = merge._when_matched_update_all_condition
+        params["when_not_matched_insert_all"] = str(
+            merge._when_not_matched_insert_all
+        ).lower()
+        params["when_not_matched_by_source_delete"] = str(
+            merge._when_not_matched_by_source_delete
+        ).lower()
+        if merge._when_not_matched_by_source_condition is not None:
+            params[
+                "when_not_matched_by_source_delete_filt"
+            ] = merge._when_not_matched_by_source_condition
+
+        self._conn._client.post(
+            f"/v1/table/{self._name}/merge_insert/",
+            data=payload,
+            params=params,
+            content_type=ARROW_STREAM_CONTENT_TYPE,
+        )
+
     def delete(self, predicate: str):
         """Delete rows from the table.
 
@@ -355,6 +440,25 @@ class RemoteTable(Table):
         payload = {"predicate": where, "updates": updates}
         self._conn._client.post(f"/v1/table/{self._name}/update/", data=payload)
 
+    def cleanup_old_versions(self, *_):
+        """cleanup_old_versions() is not supported on the LanceDB cloud"""
+        raise NotImplementedError(
+            "cleanup_old_versions() is not supported on the LanceDB cloud"
+        )
+
+    def compact_files(self, *_):
+        """compact_files() is not supported on the LanceDB cloud"""
+        raise NotImplementedError(
+            "compact_files() is not supported on the LanceDB cloud"
+        )
+
+    def count_rows(self, filter: Optional[str] = None) -> int:
+        # payload = {"filter": filter}
+        # self._conn._client.post(f"/v1/table/{self._name}/count_rows/", data=payload)
+        return NotImplementedError(
+            "count_rows() is not yet supported on the LanceDB cloud"
+        )
+
 
 def add_index(tbl: pa.Table, i: int) -> pa.Table:
     return tbl.add_column(

python/lancedb/rerankers/__init__.py

@@ -0,0 +1,15 @@
+from .base import Reranker
+from .cohere import CohereReranker
+from .colbert import ColbertReranker
+from .cross_encoder import CrossEncoderReranker
+from .linear_combination import LinearCombinationReranker
+from .openai import OpenaiReranker
+
+__all__ = [
+    "Reranker",
+    "CrossEncoderReranker",
+    "CohereReranker",
+    "LinearCombinationReranker",
+    "OpenaiReranker",
+    "ColbertReranker",
+]

python/lancedb/rerankers/base.py

@@ -0,0 +1,75 @@
+from abc import ABC, abstractmethod
+
+import numpy as np
+import pyarrow as pa
+
+
+class Reranker(ABC):
+    def __init__(self, return_score: str = "relevance"):
+        """
+        Interface for a reranker. A reranker is used to rerank the results from a
+        vector and FTS search. This is useful for combining the results from both
+        search methods.
+
+        Parameters
+        ----------
+        return_score : str, default "relevance"
+            opntions are "relevance" or "all"
+            The type of score to return. If "relevance", will return only the relevance
+            score. If "all", will return all scores from the vector and FTS search along
+            with the relevance score.
+
+        """
+        if return_score not in ["relevance", "all"]:
+            raise ValueError("score must be either 'relevance' or 'all'")
+        self.score = return_score
+
+    @abstractmethod
+    def rerank_hybrid(
+        query: str,
+        vector_results: pa.Table,
+        fts_results: pa.Table,
+    ):
+        """
+        Rerank function receives the individual results from the vector and FTS search
+        results. You can choose to use any of the results to generate the final results,
+        allowing maximum flexibility. This is mandatory to implement
+
+        Parameters
+        ----------
+        query : str
+            The input query
+        vector_results : pa.Table
+            The results from the vector search
+        fts_results : pa.Table
+            The results from the FTS search
+        """
+        pass
+
+    def merge_results(self, vector_results: pa.Table, fts_results: pa.Table):
+        """
+        Merge the results from the vector and FTS search. This is a vanilla merging
+        function that just concatenates the results and removes the duplicates.
+
+        NOTE: This doesn't take score into account. It'll keep the instance that was
+        encountered first. This is designed for rerankers that don't use the score.
+        In case you want to use the score, or support `return_scores="all"` you'll
+        have to implement your own merging function.
+
+        Parameters
+        ----------
+        vector_results : pa.Table
+            The results from the vector search
+        fts_results : pa.Table
+            The results from the FTS search
+        """
+        combined = pa.concat_tables([vector_results, fts_results], promote=True)
+        row_id = combined.column("_rowid")
+
+        # deduplicate
+        mask = np.full((combined.shape[0]), False)
+        _, mask_indices = np.unique(np.array(row_id), return_index=True)
+        mask[mask_indices] = True
+        combined = combined.filter(mask=mask)
+
+        return combined

python/lancedb/rerankers/cohere.py

@@ -0,0 +1,81 @@
+import os
+from functools import cached_property
+from typing import Union
+
+import pyarrow as pa
+
+from ..util import attempt_import_or_raise
+from .base import Reranker
+
+
+class CohereReranker(Reranker):
+    """
+    Reranks the results using the Cohere Rerank API.
+    https://docs.cohere.com/docs/rerank-guide
+
+    Parameters
+    ----------
+    model_name : str, default "rerank-english-v2.0"
+        The name of the cross encoder model to use. Available cohere models are:
+        - rerank-english-v2.0
+        - rerank-multilingual-v2.0
+    column : str, default "text"
+        The name of the column to use as input to the cross encoder model.
+    top_n : str, default None
+        The number of results to return. If None, will return all results.
+    """
+
+    def __init__(
+        self,
+        model_name: str = "rerank-english-v2.0",
+        column: str = "text",
+        top_n: Union[int, None] = None,
+        return_score="relevance",
+        api_key: Union[str, None] = None,
+    ):
+        super().__init__(return_score)
+        self.model_name = model_name
+        self.column = column
+        self.top_n = top_n
+        self.api_key = api_key
+
+    @cached_property
+    def _client(self):
+        cohere = attempt_import_or_raise("cohere")
+        if os.environ.get("COHERE_API_KEY") is None and self.api_key is None:
+            raise ValueError(
+                "COHERE_API_KEY not set. Either set it in your environment or \
+                pass it as `api_key` argument to the CohereReranker."
+            )
+        return cohere.Client(os.environ.get("COHERE_API_KEY") or self.api_key)
+
+    def rerank_hybrid(
+        self,
+        query: str,
+        vector_results: pa.Table,
+        fts_results: pa.Table,
+    ):
+        combined_results = self.merge_results(vector_results, fts_results)
+        docs = combined_results[self.column].to_pylist()
+        results = self._client.rerank(
+            query=query,
+            documents=docs,
+            top_n=self.top_n,
+            model=self.model_name,
+        )  # returns list (text, idx, relevance) attributes sorted descending by score
+        indices, scores = list(
+            zip(*[(result.index, result.relevance_score) for result in results])
+        )  # tuples
+        combined_results = combined_results.take(list(indices))
+        # add the scores
+        combined_results = combined_results.append_column(
+            "_relevance_score", pa.array(scores, type=pa.float32())
+        )
+
+        if self.score == "relevance":
+            combined_results = combined_results.drop_columns(["score", "_distance"])
+        elif self.score == "all":
+            raise NotImplementedError(
+                "return_score='all' not implemented for cohere reranker"
+            )
+        return combined_results

python/lancedb/rerankers/colbert.py

@@ -0,0 +1,109 @@
+from functools import cached_property
+
+import pyarrow as pa
+
+from ..util import attempt_import_or_raise
+from .base import Reranker
+
+
+class ColbertReranker(Reranker):
+    """
+    Reranks the results using the ColBERT model.
+
+    Parameters
+    ----------
+    model_name : str, default "colbert-ir/colbertv2.0"
+        The name of the cross encoder model to use.
+    column : str, default "text"
+        The name of the column to use as input to the cross encoder model.
+    return_score : str, default "relevance"
+        options are "relevance" or "all". Only "relevance" is supported for now.
+    """
+
+    def __init__(
+        self,
+        model_name: str = "colbert-ir/colbertv2.0",
+        column: str = "text",
+        return_score="relevance",
+    ):
+        super().__init__(return_score)
+        self.model_name = model_name
+        self.column = column
+        self.torch = attempt_import_or_raise(
+            "torch"
+        )  # import here for faster ops later
+
+    def rerank_hybrid(
+        self,
+        query: str,
+        vector_results: pa.Table,
+        fts_results: pa.Table,
+    ):
+        combined_results = self.merge_results(vector_results, fts_results)
+        docs = combined_results[self.column].to_pylist()
+
+        tokenizer, model = self._model
+
+        # Encode the query
+        query_encoding = tokenizer(query, return_tensors="pt")
+        query_embedding = model(**query_encoding).last_hidden_state.mean(dim=1)
+        scores = []
+        # Get score for each document
+        for document in docs:
+            document_encoding = tokenizer(
+                document, return_tensors="pt", truncation=True, max_length=512
+            )
+            document_embedding = model(**document_encoding).last_hidden_state
+            # Calculate MaxSim score
+            score = self.maxsim(query_embedding.unsqueeze(0), document_embedding)
+            scores.append(score.item())
+
+        # replace the self.column column with the docs
+        combined_results = combined_results.drop(self.column)
+        combined_results = combined_results.append_column(
+            self.column, pa.array(docs, type=pa.string())
+        )
+        # add the scores
+        combined_results = combined_results.append_column(
+            "_relevance_score", pa.array(scores, type=pa.float32())
+        )
+        if self.score == "relevance":
+            combined_results = combined_results.drop_columns(["score", "_distance"])
+        elif self.score == "all":
+            raise NotImplementedError(
+                "OpenAI Reranker does not support score='all' yet"
+            )
+
+        combined_results = combined_results.sort_by(
+            [("_relevance_score", "descending")]
+        )
+
+        return combined_results
+
+    @cached_property
+    def _model(self):
+        transformers = attempt_import_or_raise("transformers")
+        tokenizer = transformers.AutoTokenizer.from_pretrained(self.model_name)
+        model = transformers.AutoModel.from_pretrained(self.model_name)
+
+        return tokenizer, model
+
+    def maxsim(self, query_embedding, document_embedding):
+        # Expand dimensions for broadcasting
+        # Query: [batch, length, size] -> [batch, query, 1, size]
+        # Document: [batch, length, size] -> [batch, 1, length, size]
+        expanded_query = query_embedding.unsqueeze(2)
+        expanded_doc = document_embedding.unsqueeze(1)
+
+        # Compute cosine similarity across the embedding dimension
+        sim_matrix = self.torch.nn.functional.cosine_similarity(
+            expanded_query, expanded_doc, dim=-1
+        )
+
+        # Take the maximum similarity for each query token (across all document tokens)
+        # sim_matrix shape: [batch_size, query_length, doc_length]
+        max_sim_scores, _ = self.torch.max(sim_matrix, dim=2)
+
+        # Average these maximum scores across all query tokens
+        avg_max_sim = self.torch.mean(max_sim_scores, dim=1)
+        return avg_max_sim

python/lancedb/rerankers/cross_encoder.py

@@ -0,0 +1,74 @@
+from functools import cached_property
+from typing import Union
+
+import pyarrow as pa
+
+from ..util import attempt_import_or_raise
+from .base import Reranker
+
+
+class CrossEncoderReranker(Reranker):
+    """
+    Reranks the results using a cross encoder model. The cross encoder model is
+    used to score the query and each result. The results are then sorted by the score.
+
+    Parameters
+    ----------
+    model : str, default "cross-encoder/ms-marco-TinyBERT-L-6"
+        The name of the cross encoder model to use. See the sentence transformers
+        documentation for a list of available models.
+    column : str, default "text"
+        The name of the column to use as input to the cross encoder model.
+    device : str, default None
+        The device to use for the cross encoder model. If None, will use "cuda"
+        if available, otherwise "cpu".
+    """
+
+    def __init__(
+        self,
+        model_name: str = "cross-encoder/ms-marco-TinyBERT-L-6",
+        column: str = "text",
+        device: Union[str, None] = None,
+        return_score="relevance",
+    ):
+        super().__init__(return_score)
+        torch = attempt_import_or_raise("torch")
+        self.model_name = model_name
+        self.column = column
+        self.device = device
+        if self.device is None:
+            self.device = "cuda" if torch.cuda.is_available() else "cpu"
+
+    @cached_property
+    def model(self):
+        sbert = attempt_import_or_raise("sentence_transformers")
+        cross_encoder = sbert.CrossEncoder(self.model_name)
+
+        return cross_encoder
+
+    def rerank_hybrid(
+        self,
+        query: str,
+        vector_results: pa.Table,
+        fts_results: pa.Table,
+    ):
+        combined_results = self.merge_results(vector_results, fts_results)
+        passages = combined_results[self.column].to_pylist()
+        cross_inp = [[query, passage] for passage in passages]
+        cross_scores = self.model.predict(cross_inp)
+        combined_results = combined_results.append_column(
+            "_relevance_score", pa.array(cross_scores, type=pa.float32())
+        )
+
+        # sort the results by _score
+        if self.score == "relevance":
+            combined_results = combined_results.drop_columns(["score", "_distance"])
+        elif self.score == "all":
+            raise NotImplementedError(
+                "return_score='all' not implemented for CrossEncoderReranker"
+            )
+        combined_results = combined_results.sort_by(
+            [("_relevance_score", "descending")]
+        )
+
+        return combined_results

python/lancedb/rerankers/linear_combination.py

@@ -0,0 +1,117 @@
+from typing import List
+
+import pyarrow as pa
+
+from .base import Reranker
+
+
+class LinearCombinationReranker(Reranker):
+    """
+    Reranks the results using a linear combination of the scores from the
+    vector and FTS search. For missing scores, fill with `fill` value.
+    Parameters
+    ----------
+    weight : float, default 0.7
+        The weight to give to the vector score. Must be between 0 and 1.
+    fill : float, default 1.0
+        The score to give to results that are only in one of the two result sets.
+        This is treated as penalty, so a higher value means a lower score.
+        TODO: We should just hardcode this--
+        its pretty confusing as we invert scores to calculate final score
+    return_score : str, default "relevance"
+        opntions are "relevance" or "all"
+        The type of score to return. If "relevance", will return only the relevance
+        score. If "all", will return all scores from the vector and FTS search along
+        with the relevance score.
+    """
+
+    def __init__(
+        self, weight: float = 0.7, fill: float = 1.0, return_score="relevance"
+    ):
+        if weight < 0 or weight > 1:
+            raise ValueError("weight must be between 0 and 1.")
+        super().__init__(return_score)
+        self.weight = weight
+        self.fill = fill
+
+    def rerank_hybrid(
+        self,
+        query: str,  # noqa: F821
+        vector_results: pa.Table,
+        fts_results: pa.Table,
+    ):
+        combined_results = self.merge_results(vector_results, fts_results, self.fill)
+
+        return combined_results
+
+    def merge_results(
+        self, vector_results: pa.Table, fts_results: pa.Table, fill: float
+    ):
+        # If both are empty then just return an empty table
+        if len(vector_results) == 0 and len(fts_results) == 0:
+            return vector_results
+        # If one is empty then return the other
+        if len(vector_results) == 0:
+            return fts_results
+        if len(fts_results) == 0:
+            return vector_results
+
+        # sort both input tables on _rowid
+        combined_list = []
+        vector_list = vector_results.sort_by("_rowid").to_pylist()
+        fts_list = fts_results.sort_by("_rowid").to_pylist()
+        i, j = 0, 0
+        while i < len(vector_list):
+            if j >= len(fts_list):
+                for vi in vector_list[i:]:
+                    vi["_relevance_score"] = self._combine_score(vi["_distance"], fill)
+                    combined_list.append(vi)
+                break
+
+            vi = vector_list[i]
+            fj = fts_list[j]
+            # invert the fts score from relevance to distance
+            inverted_fts_score = self._invert_score(fj["score"])
+            if vi["_rowid"] == fj["_rowid"]:
+                vi["_relevance_score"] = self._combine_score(
+                    vi["_distance"], inverted_fts_score
+                )
+                vi["score"] = fj["score"]  # keep the original score
+                combined_list.append(vi)
+                i += 1
+                j += 1
+            elif vector_list[i]["_rowid"] < fts_list[j]["_rowid"]:
+                vi["_relevance_score"] = self._combine_score(vi["_distance"], fill)
+                combined_list.append(vi)
+                i += 1
+            else:
+                fj["_relevance_score"] = self._combine_score(inverted_fts_score, fill)
+                combined_list.append(fj)
+                j += 1
+        if j < len(fts_list) - 1:
+            for fj in fts_list[j:]:
+                fj["_relevance_score"] = self._combine_score(inverted_fts_score, fill)
+                combined_list.append(fj)
+
+        relevance_score_schema = pa.schema(
+            [
+                pa.field("_relevance_score", pa.float32()),
+            ]
+        )
+        combined_schema = pa.unify_schemas(
+            [vector_results.schema, fts_results.schema, relevance_score_schema]
+        )
+        tbl = pa.Table.from_pylist(combined_list, schema=combined_schema).sort_by(
+            [("_relevance_score", "descending")]
+        )
+        if self.score == "relevance":
+            tbl = tbl.drop_columns(["score", "_distance"])
+        return tbl
+
+    def _combine_score(self, score1, score2):
+        # these scores represent distance
+        return 1 - (self.weight * score1 + (1 - self.weight) * score2)
+
+    def _invert_score(self, scores: List[float]):
+        # Invert the scores between relevance and distance
+        return 1 - scores

python/lancedb/rerankers/openai.py

@@ -0,0 +1,104 @@
+import json
+import os
+from functools import cached_property
+from typing import Optional
+
+import pyarrow as pa
+
+from ..util import attempt_import_or_raise
+from .base import Reranker
+
+
+class OpenaiReranker(Reranker):
+    """
+    Reranks the results using the OpenAI API.
+    WARNING: This is a prompt based reranker that uses chat model that is
+    not a dedicated reranker API. This should be treated as experimental.
+
+    Parameters
+    ----------
+    model_name : str, default "gpt-4-turbo-preview"
+        The name of the cross encoder model to use.
+    column : str, default "text"
+        The name of the column to use as input to the cross encoder model.
+    return_score : str, default "relevance"
+        options are "relevance" or "all". Only "relevance" is supported for now.
+    api_key : str, default None
+        The API key to use. If None, will use the OPENAI_API_KEY environment variable.
+    """
+
+    def __init__(
+        self,
+        model_name: str = "gpt-4-turbo-preview",
+        column: str = "text",
+        return_score="relevance",
+        api_key: Optional[str] = None,
+    ):
+        super().__init__(return_score)
+        self.model_name = model_name
+        self.column = column
+        self.api_key = api_key
+
+    def rerank_hybrid(
+        self,
+        query: str,
+        vector_results: pa.Table,
+        fts_results: pa.Table,
+    ):
+        combined_results = self.merge_results(vector_results, fts_results)
+        docs = combined_results[self.column].to_pylist()
+        response = self._client.chat.completions.create(
+            model=self.model_name,
+            response_format={"type": "json_object"},
+            temperature=0,
+            messages=[
+                {
+                    "role": "system",
+                    "content": "You are an expert relevance ranker. Given a list of\
+                        documents and a query, your job is to determine the relevance\
+                        each document is for answering the query. Your output is JSON,\
+                        which is a list of documents. Each document has two fields,\
+                        content and relevance_score.  relevance_score is from 0.0 to\
+                        1.0 indicating the relevance of the text to the given query.\
+                        Make sure to include all documents in the response.",
+                },
+                {"role": "user", "content": f"Query: {query} Docs: {docs}"},
+            ],
+        )
+        results = json.loads(response.choices[0].message.content)["documents"]
+        docs, scores = list(
+            zip(*[(result["content"], result["relevance_score"]) for result in results])
+        )  # tuples
+        # replace the self.column column with the docs
+        combined_results = combined_results.drop(self.column)
+        combined_results = combined_results.append_column(
+            self.column, pa.array(docs, type=pa.string())
+        )
+        # add the scores
+        combined_results = combined_results.append_column(
+            "_relevance_score", pa.array(scores, type=pa.float32())
+        )
+        if self.score == "relevance":
+            combined_results = combined_results.drop_columns(["score", "_distance"])
+        elif self.score == "all":
+            raise NotImplementedError(
+                "OpenAI Reranker does not support score='all' yet"
+            )
+
+        combined_results = combined_results.sort_by(
+            [("_relevance_score", "descending")]
+        )
+
+        return combined_results
+
+    @cached_property
+    def _client(self):
+        openai = attempt_import_or_raise(
+            "openai"
+        )  # TODO: force version or handle versions < 1.0
+        if os.environ.get("OPENAI_API_KEY") is None and self.api_key is None:
+            raise ValueError(
+                "OPENAI_API_KEY not set. Either set it in your environment or \
+                pass it as `api_key` argument to the CohereReranker."
+            )
+        return openai.OpenAI(api_key=os.environ.get("OPENAI_API_KEY") or self.api_key)

python/lancedb/table.py

@@ -14,9 +14,12 @@
 from __future__ import annotations
 
 import inspect
+import time
 from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from datetime import timedelta
 from functools import cached_property
-from typing import TYPE_CHECKING, Any, Dict, Iterable, List, Optional, Union
+from typing import TYPE_CHECKING, Any, Dict, Iterable, List, Optional, Tuple, Union
 
 import lance
 import numpy as np
@@ -28,10 +31,12 @@ from lance.vector import vec_to_table
 
 from .common import DATA, VEC, VECTOR_COLUMN_NAME
 from .embeddings import EmbeddingFunctionConfig, EmbeddingFunctionRegistry
+from .merge import LanceMergeInsertBuilder
 from .pydantic import LanceModel, model_to_dict
 from .query import LanceQueryBuilder, Query
 from .util import (
     fs_from_uri,
+    inf_vector_column_query,
     join_uri,
     safe_import_pandas,
     safe_import_polars,
@@ -40,8 +45,6 @@ from .util import (
 from .utils.events import register_event
 
 if TYPE_CHECKING:
-    from datetime import timedelta
-
     import PIL
     from lance.dataset import CleanupStats, ReaderLike
 
@@ -175,6 +178,18 @@ class Table(ABC):
         """
         raise NotImplementedError
 
+    @abstractmethod
+    def count_rows(self, filter: Optional[str] = None) -> int:
+        """
+        Count the number of rows in the table.
+
+        Parameters
+        ----------
+        filter: str, optional
+            A SQL where clause to filter the rows to count.
+        """
+        raise NotImplementedError
+
     def to_pandas(self) -> "pd.DataFrame":
         """Return the table as a pandas DataFrame.
 
@@ -298,7 +313,7 @@ class Table(ABC):
 
             import lance
 
-            dataset = lance.dataset("/tmp/images.lance")
+            dataset = lance.dataset("./images.lance")
             dataset.create_scalar_index("category")
         """
         raise NotImplementedError
@@ -335,11 +350,71 @@ class Table(ABC):
         """
         raise NotImplementedError
 
+    def merge_insert(self, on: Union[str, Iterable[str]]) -> LanceMergeInsertBuilder:
+        """
+        Returns a [`LanceMergeInsertBuilder`][lancedb.merge.LanceMergeInsertBuilder]
+        that can be used to create a "merge insert" operation
+
+        This operation can add rows, update rows, and remove rows all in a single
+        transaction. It is a very generic tool that can be used to create
+        behaviors like "insert if not exists", "update or insert (i.e. upsert)",
+        or even replace a portion of existing data with new data (e.g. replace
+        all data where month="january")
+
+        The merge insert operation works by combining new data from a
+        **source table** with existing data in a **target table** by using a
+        join.  There are three categories of records.
+
+        "Matched" records are records that exist in both the source table and
+        the target table. "Not matched" records exist only in the source table
+        (e.g. these are new data) "Not matched by source" records exist only
+        in the target table (this is old data)
+
+        The builder returned by this method can be used to customize what
+        should happen for each category of data.
+
+        Please note that the data may appear to be reordered as part of this
+        operation.  This is because updated rows will be deleted from the
+        dataset and then reinserted at the end with the new values.
+
+        Parameters
+        ----------
+
+        on: Union[str, Iterable[str]]
+            A column (or columns) to join on.  This is how records from the
+            source table and target table are matched.  Typically this is some
+            kind of key or id column.
+
+        Examples
+        --------
+        >>> import lancedb
+        >>> data = pa.table({"a": [2, 1, 3], "b": ["a", "b", "c"]})
+        >>> db = lancedb.connect("./.lancedb")
+        >>> table = db.create_table("my_table", data)
+        >>> new_data = pa.table({"a": [2, 3, 4], "b": ["x", "y", "z"]})
+        >>> # Perform a "upsert" operation
+        >>> table.merge_insert("a")             \\
+        ...      .when_matched_update_all()     \\
+        ...      .when_not_matched_insert_all() \\
+        ...      .execute(new_data)
+        >>> # The order of new rows is non-deterministic since we use
+        >>> # a hash-join as part of this operation and so we sort here
+        >>> table.to_arrow().sort_by("a").to_pandas()
+           a  b
+        0  1  b
+        1  2  x
+        2  3  y
+        3  4  z
+        """
+        on = [on] if isinstance(on, str) else list(on.iter())
+
+        return LanceMergeInsertBuilder(self, on)
+
     @abstractmethod
     def search(
         self,
-        query: Optional[Union[VEC, str, "PIL.Image.Image"]] = None,
-        vector_column_name: str = VECTOR_COLUMN_NAME,
+        query: Optional[Union[VEC, str, "PIL.Image.Image", Tuple]] = None,
+        vector_column_name: Optional[str] = None,
         query_type: str = "auto",
     ) -> LanceQueryBuilder:
         """Create a search query to find the nearest neighbors
@@ -359,7 +434,7 @@ class Table(ABC):
         ... ]
         >>> table = db.create_table("my_table", data)
         >>> query = [0.4, 1.4, 2.4]
-        >>> (table.search(query, vector_column_name="vector")
+        >>> (table.search(query)
         ...     .where("original_width > 1000", prefilter=True)
         ...     .select(["caption", "original_width"])
         ...     .limit(2)
@@ -378,12 +453,19 @@ class Table(ABC):
 
             - If None then the select/where/limit clauses are applied to filter
             the table
-        vector_column_name: str
+        vector_column_name: str, optional
             The name of the vector column to search.
-            *default "vector"*
+
+            The vector column needs to be a pyarrow fixed size list type
+
+            - If not specified then the vector column is inferred from
+            the table schema
+
+            - If the table has multiple vector columns then the *vector_column_name*
+            needs to be specified. Otherwise, an error is raised.
         query_type: str
             *default "auto"*.
-            Acceptable types are: "vector", "fts", or "auto"
+            Acceptable types are: "vector", "fts", "hybrid", or "auto"
 
             - If "auto" then the query type is inferred from the query;
 
@@ -415,6 +497,16 @@ class Table(ABC):
     def _execute_query(self, query: Query) -> pa.Table:
         pass
 
+    @abstractmethod
+    def _do_merge(
+        self,
+        merge: LanceMergeInsertBuilder,
+        new_data: DATA,
+        on_bad_vectors: str,
+        fill_value: float,
+    ):
+        pass
+
     @abstractmethod
     def delete(self, where: str):
         """Delete rows from the table.
@@ -522,24 +614,192 @@ class Table(ABC):
         """
         raise NotImplementedError
 
+    @abstractmethod
+    def cleanup_old_versions(
+        self,
+        older_than: Optional[timedelta] = None,
+        *,
+        delete_unverified: bool = False,
+    ) -> CleanupStats:
+        """
+        Clean up old versions of the table, freeing disk space.
+
+        Note: This function is not available in LanceDb Cloud (since LanceDb
+        Cloud manages cleanup for you automatically)
+
+        Parameters
+        ----------
+        older_than: timedelta, default None
+            The minimum age of the version to delete. If None, then this defaults
+            to two weeks.
+        delete_unverified: bool, default False
+            Because they may be part of an in-progress transaction, files newer
+            than 7 days old are not deleted by default. If you are sure that
+            there are no in-progress transactions, then you can set this to True
+            to delete all files older than `older_than`.
+
+        Returns
+        -------
+        CleanupStats
+            The stats of the cleanup operation, including how many bytes were
+            freed.
+        """
+
+    @abstractmethod
+    def compact_files(self, *args, **kwargs):
+        """
+        Run the compaction process on the table.
+
+        Note: This function is not available in LanceDb Cloud (since LanceDb
+        Cloud manages compaction for you automatically)
+
+        This can be run after making several small appends to optimize the table
+        for faster reads.
+
+        Arguments are passed onto :meth:`lance.dataset.DatasetOptimizer.compact_files`.
+        For most cases, the default should be fine.
+        """
+
+
+class _LanceDatasetRef(ABC):
+    @property
+    @abstractmethod
+    def dataset(self) -> LanceDataset:
+        pass
+
+    @property
+    @abstractmethod
+    def dataset_mut(self) -> LanceDataset:
+        pass
+
+
+@dataclass
+class _LanceLatestDatasetRef(_LanceDatasetRef):
+    """Reference to the latest version of a LanceDataset."""
+
+    uri: str
+    read_consistency_interval: Optional[timedelta] = None
+    last_consistency_check: Optional[float] = None
+    _dataset: Optional[LanceDataset] = None
+
+    @property
+    def dataset(self) -> LanceDataset:
+        if not self._dataset:
+            self._dataset = lance.dataset(self.uri)
+            self.last_consistency_check = time.monotonic()
+        elif self.read_consistency_interval is not None:
+            now = time.monotonic()
+            diff = timedelta(seconds=now - self.last_consistency_check)
+            if (
+                self.last_consistency_check is None
+                or diff > self.read_consistency_interval
+            ):
+                self._dataset = self._dataset.checkout_version(
+                    self._dataset.latest_version
+                )
+                self.last_consistency_check = time.monotonic()
+        return self._dataset
+
+    @dataset.setter
+    def dataset(self, value: LanceDataset):
+        self._dataset = value
+        self.last_consistency_check = time.monotonic()
+
+    @property
+    def dataset_mut(self) -> LanceDataset:
+        return self.dataset
+
+
+@dataclass
+class _LanceTimeTravelRef(_LanceDatasetRef):
+    uri: str
+    version: int
+    _dataset: Optional[LanceDataset] = None
+
+    @property
+    def dataset(self) -> LanceDataset:
+        if not self._dataset:
+            self._dataset = lance.dataset(self.uri, version=self.version)
+        return self._dataset
+
+    @dataset.setter
+    def dataset(self, value: LanceDataset):
+        self._dataset = value
+        self.version = value.version
+
+    @property
+    def dataset_mut(self) -> LanceDataset:
+        raise ValueError(
+            "Cannot mutate table reference fixed at version "
+            f"{self.version}. Call checkout_latest() to get a mutable "
+            "table reference."
+        )
+
 
 class LanceTable(Table):
     """
     A table in a LanceDB database.
+
+    This can be opened in two modes: standard and time-travel.
+
+    Standard mode is the default. In this mode, the table is mutable and tracks
+    the latest version of the table. The level of read consistency is controlled
+    by the `read_consistency_interval` parameter on the connection.
+
+    Time-travel mode is activated by specifying a version number. In this mode,
+    the table is immutable and fixed to a specific version. This is useful for
+    querying historical versions of the table.
     """
 
-    def __init__(self, connection: "LanceDBConnection", name: str, version: int = None):
+    def __init__(
+        self,
+        connection: "LanceDBConnection",
+        name: str,
+        version: Optional[int] = None,
+    ):
         self._conn = connection
         self.name = name
-        self._version = version
 
-    def _reset_dataset(self, version=None):
-        try:
-            if "_dataset" in self.__dict__:
-                del self.__dict__["_dataset"]
-            self._version = version
-        except AttributeError:
-            pass
+        if version is not None:
+            self._ref = _LanceTimeTravelRef(
+                uri=self._dataset_uri,
+                version=version,
+            )
+        else:
+            self._ref = _LanceLatestDatasetRef(
+                uri=self._dataset_uri,
+                read_consistency_interval=connection.read_consistency_interval,
+            )
+
+    @classmethod
+    def open(cls, db, name, **kwargs):
+        tbl = cls(db, name, **kwargs)
+        fs, path = fs_from_uri(tbl._dataset_uri)
+        file_info = fs.get_file_info(path)
+        if file_info.type != pa.fs.FileType.Directory:
+            raise FileNotFoundError(
+                f"Table {name} does not exist."
+                f"Please first call db.create_table({name}, data)"
+            )
+        register_event("open_table")
+
+        return tbl
+
+    @property
+    def _dataset_uri(self) -> str:
+        return join_uri(self._conn.uri, f"{self.name}.lance")
+
+    @property
+    def _dataset(self) -> LanceDataset:
+        return self._ref.dataset
+
+    @property
+    def _dataset_mut(self) -> LanceDataset:
+        return self._ref.dataset_mut
+
+    def to_lance(self) -> LanceDataset:
+        """Return the LanceDataset backing this table."""
+        return self._dataset
 
     @property
     def schema(self) -> pa.Schema:
@@ -567,6 +827,9 @@ class LanceTable(Table):
         keep writing to the dataset starting from an old version, then use
         the `restore` function.
 
+        Calling this method will set the table into time-travel mode. If you
+        wish to return to standard mode, call `checkout_latest`.
+
         Parameters
         ----------
         version : int
@@ -591,15 +854,13 @@ class LanceTable(Table):
                vector    type
         0  [1.1, 0.9]  vector
         """
-        max_ver = max([v["version"] for v in self._dataset.versions()])
+        max_ver = self._dataset.latest_version
         if version < 1 or version > max_ver:
             raise ValueError(f"Invalid version {version}")
-        self._reset_dataset(version=version)
 
         try:
-            # Accessing the property updates the cached value
-            _ = self._dataset
-        except Exception as e:
+            ds = self._dataset.checkout_version(version)
+        except IOError as e:
             if "not found" in str(e):
                 raise ValueError(
                     f"Version {version} no longer exists. Was it cleaned up?"
@@ -607,6 +868,27 @@ class LanceTable(Table):
             else:
                 raise e
 
+        self._ref = _LanceTimeTravelRef(
+            uri=self._dataset_uri,
+            version=version,
+        )
+        # We've already loaded the version so we can populate it directly.
+        self._ref.dataset = ds
+
+    def checkout_latest(self):
+        """Checkout the latest version of the table. This is an in-place operation.
+
+        The table will be set back into standard mode, and will track the latest
+        version of the table.
+        """
+        self.checkout(self._dataset.latest_version)
+        ds = self._ref.dataset
+        self._ref = _LanceLatestDatasetRef(
+            uri=self._dataset_uri,
+            read_consistency_interval=self._conn.read_consistency_interval,
+        )
+        self._ref.dataset = ds
+
     def restore(self, version: int = None):
         """Restore a version of the table. This is an in-place operation.
 
@@ -641,7 +923,7 @@ class LanceTable(Table):
         >>> len(table.list_versions())
         4
         """
-        max_ver = max([v["version"] for v in self._dataset.versions()])
+        max_ver = self._dataset.latest_version
         if version is None:
             version = self.version
         elif version < 1 or version > max_ver:
@@ -649,29 +931,30 @@ class LanceTable(Table):
         else:
             self.checkout(version)
 
-        if version == max_ver:
-            # no-op if restoring the latest version
-            return
+        ds = self._dataset
 
-        self._dataset.restore()
-        self._reset_dataset()
+        # no-op if restoring the latest version
+        if version != max_ver:
+            ds.restore()
+
+        self._ref = _LanceLatestDatasetRef(
+            uri=self._dataset_uri,
+            read_consistency_interval=self._conn.read_consistency_interval,
+        )
+        self._ref.dataset = ds
 
     def count_rows(self, filter: Optional[str] = None) -> int:
-        """
-        Count the number of rows in the table.
-
-        Parameters
-        ----------
-        filter: str, optional
-            A SQL where clause to filter the rows to count.
-        """
         return self._dataset.count_rows(filter)
 
     def __len__(self):
         return self.count_rows()
 
     def __repr__(self) -> str:
-        return f"LanceTable({self.name})"
+        val = f'{self.__class__.__name__}(connection={self._conn!r}, name="{self.name}"'
+        if isinstance(self._ref, _LanceTimeTravelRef):
+            val += f", version={self._ref.version}"
+        val += ")"
+        return val
 
     def __str__(self) -> str:
         return self.__repr__()
@@ -721,10 +1004,6 @@ class LanceTable(Table):
             self.to_lance(), allow_pyarrow_filter=False, batch_size=batch_size
         )
 
-    @property
-    def _dataset_uri(self) -> str:
-        return join_uri(self._conn.uri, f"{self.name}.lance")
-
     def create_index(
         self,
         metric="L2",
@@ -736,7 +1015,7 @@ class LanceTable(Table):
         index_cache_size: Optional[int] = None,
     ):
         """Create an index on the table."""
-        self._dataset.create_index(
+        self._dataset_mut.create_index(
             column=vector_column_name,
             index_type="IVF_PQ",
             metric=metric,
@@ -746,11 +1025,12 @@ class LanceTable(Table):
             accelerator=accelerator,
             index_cache_size=index_cache_size,
         )
-        self._reset_dataset()
         register_event("create_index")
 
     def create_scalar_index(self, column: str, *, replace: bool = True):
-        self._dataset.create_scalar_index(column, index_type="BTREE", replace=replace)
+        self._dataset_mut.create_scalar_index(
+            column, index_type="BTREE", replace=replace
+        )
 
     def create_fts_index(
         self,
@@ -793,14 +1073,6 @@ class LanceTable(Table):
     def _get_fts_index_path(self):
         return join_uri(self._dataset_uri, "_indices", "tantivy")
 
-    @cached_property
-    def _dataset(self) -> LanceDataset:
-        return lance.dataset(self._dataset_uri, version=self._version)
-
-    def to_lance(self) -> LanceDataset:
-        """Return the LanceDataset backing this table."""
-        return self._dataset
-
     def add(
         self,
         data: DATA,
@@ -839,8 +1111,11 @@ class LanceTable(Table):
             on_bad_vectors=on_bad_vectors,
             fill_value=fill_value,
         )
-        lance.write_dataset(data, self._dataset_uri, schema=self.schema, mode=mode)
-        self._reset_dataset()
+        # Access the dataset_mut property to ensure that the dataset is mutable.
+        self._ref.dataset_mut
+        self._ref.dataset = lance.write_dataset(
+            data, self._dataset_uri, schema=self.schema, mode=mode
+        )
         register_event("add")
 
     def merge(
@@ -901,10 +1176,9 @@ class LanceTable(Table):
             other_table = other_table.to_lance()
         if isinstance(other_table, LanceDataset):
             other_table = other_table.to_table()
-        self._dataset.merge(
+        self._ref.dataset = self._dataset_mut.merge(
             other_table, left_on=left_on, right_on=right_on, schema=schema
         )
-        self._reset_dataset()
         register_event("merge")
 
     @cached_property
@@ -924,8 +1198,8 @@ class LanceTable(Table):
 
     def search(
         self,
-        query: Optional[Union[VEC, str, "PIL.Image.Image"]] = None,
-        vector_column_name: str = VECTOR_COLUMN_NAME,
+        query: Optional[Union[VEC, str, "PIL.Image.Image", Tuple]] = None,
+        vector_column_name: Optional[str] = None,
         query_type: str = "auto",
     ) -> LanceQueryBuilder:
         """Create a search query to find the nearest neighbors
@@ -943,7 +1217,7 @@ class LanceTable(Table):
         ... ]
         >>> table = db.create_table("my_table", data)
         >>> query = [0.4, 1.4, 2.4]
-        >>> (table.search(query, vector_column_name="vector")
+        >>> (table.search(query)
         ...     .where("original_width > 1000", prefilter=True)
         ...     .select(["caption", "original_width"])
         ...     .limit(2)
@@ -962,8 +1236,17 @@ class LanceTable(Table):
 
             - If None then the select/[where][sql]/limit clauses are applied
             to filter the table
-        vector_column_name: str, default "vector"
+        vector_column_name: str, optional
             The name of the vector column to search.
+
+            The vector column needs to be a pyarrow fixed size list type
+            *default "vector"*
+
+            - If not specified then the vector column is inferred from
+            the table schema
+
+            - If the table has multiple vector columns then the *vector_column_name*
+            needs to be specified. Otherwise, an error is raised.
         query_type: str, default "auto"
             "vector", "fts", or "auto"
             If "auto" then the query type is inferred from the query;
@@ -981,6 +1264,8 @@ class LanceTable(Table):
             and also the "_distance" column which is the distance between the query
             vector and the returned vector.
         """
+        if vector_column_name is None and query is not None:
+            vector_column_name = inf_vector_column_query(self.schema)
         register_event("search_table")
         return LanceQueryBuilder.create(
             self, query, query_type, vector_column_name=vector_column_name
@@ -1107,22 +1392,8 @@ class LanceTable(Table):
         register_event("create_table")
         return new_table
 
-    @classmethod
-    def open(cls, db, name):
-        tbl = cls(db, name)
-        fs, path = fs_from_uri(tbl._dataset_uri)
-        file_info = fs.get_file_info(path)
-        if file_info.type != pa.fs.FileType.Directory:
-            raise FileNotFoundError(
-                f"Table {name} does not exist."
-                f"Please first call db.create_table({name}, data)"
-            )
-        register_event("open_table")
-
-        return tbl
-
     def delete(self, where: str):
-        self._dataset.delete(where)
+        self._dataset_mut.delete(where)
 
     def update(
         self,
@@ -1176,12 +1447,12 @@ class LanceTable(Table):
         if values is not None:
             values_sql = {k: value_to_sql(v) for k, v in values.items()}
 
-        self.to_lance().update(values_sql, where)
-        self._reset_dataset()
+        self._dataset_mut.update(values_sql, where)
         register_event("update")
 
     def _execute_query(self, query: Query) -> pa.Table:
         ds = self.to_lance()
+
         return ds.to_table(
             columns=query.columns,
             filter=query.filter,
@@ -1194,8 +1465,34 @@ class LanceTable(Table):
                 "nprobes": query.nprobes,
                 "refine_factor": query.refine_factor,
             },
+            with_row_id=query.with_row_id,
         )
 
+    def _do_merge(
+        self,
+        merge: LanceMergeInsertBuilder,
+        new_data: DATA,
+        on_bad_vectors: str,
+        fill_value: float,
+    ):
+        new_data = _sanitize_data(
+            new_data,
+            self.schema,
+            metadata=self.schema.metadata,
+            on_bad_vectors=on_bad_vectors,
+            fill_value=fill_value,
+        )
+        ds = self.to_lance()
+        builder = ds.merge_insert(merge._on)
+        if merge._when_matched_update_all:
+            builder.when_matched_update_all(merge._when_matched_update_all_condition)
+        if merge._when_not_matched_insert_all:
+            builder.when_not_matched_insert_all()
+        if merge._when_not_matched_by_source_delete:
+            cond = merge._when_not_matched_by_source_condition
+            builder.when_not_matched_by_source_delete(cond)
+        builder.execute(new_data)
+
     def cleanup_old_versions(
         self,
         older_than: Optional[timedelta] = None,
@@ -1233,8 +1530,9 @@ class LanceTable(Table):
         This can be run after making several small appends to optimize the table
         for faster reads.
 
-        Arguments are passed onto :meth:`lance.dataset.DatasetOptimizer.compact_files`.
-        For most cases, the default should be fine.
+        Arguments are passed onto `lance.dataset.DatasetOptimizer.compact_files`.
+         (see Lance documentation for more details) For most cases, the default
+        should be fine.
         """
         return self.to_lance().optimize.compact_files(*args, **kwargs)
 

python/lancedb/util.py

@@ -11,6 +11,7 @@
 #  See the License for the specific language governing permissions and
 #  limitations under the License.
 
+import importlib
 import os
 import pathlib
 from datetime import date, datetime
@@ -19,6 +20,7 @@ from typing import Tuple, Union
 from urllib.parse import urlparse
 
 import numpy as np
+import pyarrow as pa
 import pyarrow.fs as pa_fs
 
 
@@ -114,6 +116,25 @@ def join_uri(base: Union[str, pathlib.Path], *parts: str) -> str:
     return "/".join([p.rstrip("/") for p in [base, *parts]])
 
 
+def attempt_import_or_raise(module: str, mitigation=None):
+    """
+    Import the specified module. If the module is not installed,
+    raise an ImportError with a helpful message.
+
+    Parameters
+    ----------
+    module : str
+        The name of the module to import
+    mitigation : Optional[str]
+        The package(s) to install to mitigate the error.
+        If not provided then the module name will be used.
+    """
+    try:
+        return importlib.import_module(module)
+    except ImportError:
+        raise ImportError(f"Please install {mitigation or module}")
+
+
 def safe_import_pandas():
     try:
         import pandas as pd
@@ -132,6 +153,44 @@ def safe_import_polars():
         return None
 
 
+def inf_vector_column_query(schema: pa.Schema) -> str:
+    """
+    Get the vector column name
+
+    Parameters
+    ----------
+    schema : pa.Schema
+        The schema of the vector column.
+
+    Returns
+    -------
+    str: the vector column name.
+    """
+    vector_col_name = ""
+    vector_col_count = 0
+    for field_name in schema.names:
+        field = schema.field(field_name)
+        if pa.types.is_fixed_size_list(field.type) and pa.types.is_floating(
+            field.type.value_type
+        ):
+            vector_col_count += 1
+            if vector_col_count > 1:
+                raise ValueError(
+                    "Schema has more than one vector column. "
+                    "Please specify the vector column name "
+                    "for vector search"
+                )
+                break
+            elif vector_col_count == 1:
+                vector_col_name = field_name
+    if vector_col_count == 0:
+        raise ValueError(
+            "There is no vector column in the data. "
+            "Please specify the vector column name for vector search"
+        )
+    return vector_col_name
+
+
 @singledispatch
 def value_to_sql(value):
     raise NotImplementedError("SQL conversion is not implemented for this type")

python/pyproject.toml

@@ -1,9 +1,9 @@
 [project]
 name = "lancedb"
-version = "0.5.1"
+version = "0.5.5"
 dependencies = [
     "deprecation",
-    "pylance==0.9.9",
+    "pylance==0.9.16",
     "ratelimiter~=1.0",
     "retry>=0.9.2",
     "tqdm>=4.27.0",
@@ -48,11 +48,12 @@ classifiers = [
 repository = "https://github.com/lancedb/lancedb"
 
 [project.optional-dependencies]
-tests = ["aiohttp", "pandas>=1.4", "pytest", "pytest-mock", "pytest-asyncio", "duckdb", "pytz", "polars"]
+tests = ["aiohttp", "pandas>=1.4", "pytest", "pytest-mock", "pytest-asyncio", "duckdb", "pytz", "polars>=0.19"]
 dev = ["ruff", "pre-commit"]
 docs = ["mkdocs", "mkdocs-jupyter", "mkdocs-material", "mkdocstrings[python]"]
 clip = ["torch", "pillow", "open-clip"]
-embeddings = ["openai>=1.6.1", "sentence-transformers", "torch", "pillow", "open-clip-torch", "cohere", "InstructorEmbedding"]
+embeddings = ["openai>=1.6.1", "sentence-transformers", "torch", "pillow", "open-clip-torch", "cohere", "huggingface_hub",
+                "InstructorEmbedding", "google.generativeai", "boto3>=1.28.57", "awscli>=1.29.57", "botocore>=1.31.57"]
 
 [project.scripts]
 lancedb = "lancedb.cli.cli:cli"
@@ -65,7 +66,8 @@ build-backend = "setuptools.build_meta"
 select = ["F", "E", "W", "I", "G", "TCH", "PERF"]
 
 [tool.pytest.ini_options]
-addopts = "--strict-markers"
+addopts = "--strict-markers --ignore-glob=lancedb/embeddings/*.py"
+
 markers = [
     "slow: marks tests as slow (deselect with '-m \"not slow\"')",
     "asyncio"

python/tests/test_embeddings.py

@@ -88,6 +88,7 @@ def test_embedding_function(tmp_path):
     assert np.allclose(actual, expected)
 
 
+@pytest.mark.slow
 def test_embedding_function_rate_limit(tmp_path):
     def _get_schema_from_model(model):
         class Schema(LanceModel):