Bump version: 0.3.8 → 0.3.9

was passing this at the wrong position

.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.3.7
+current_version = 0.3.9
 commit = True
 message = Bump version: {current_version} → {new_version}
 tag = True

Cargo.toml

@@ -5,9 +5,10 @@ exclude = ["python"]
 resolver = "2"
 
 [workspace.dependencies]
-lance = { "version" = "=0.8.14", "features" = ["dynamodb"] }
-lance-linalg = { "version" = "=0.8.14" }
-lance-testing = { "version" = "=0.8.14" }
+lance = { "version" = "=0.8.17", "features" = ["dynamodb"] }
+lance-index = { "version" = "=0.8.17" }
+lance-linalg = { "version" = "=0.8.17" }
+lance-testing = { "version" = "=0.8.17" }
 # Note that this one does not include pyarrow
 arrow = { version = "47.0.0", optional = false }
 arrow-array = "47.0"

docs/mkdocs.yml

@@ -1,4 +1,5 @@
 site_name: LanceDB Docs
+site_url: https://lancedb.github.io/lancedb/
 repo_url: https://github.com/lancedb/lancedb
 edit_uri: https://github.com/lancedb/lancedb/tree/main/docs/src
 repo_name: lancedb/lancedb
@@ -79,6 +80,7 @@ nav:
     - Ingest Embedding Functions: embeddings/embedding_functions.md
     - Available Functions: embeddings/default_embedding_functions.md
     - Create Custom Embedding Functions: embeddings/api.md
+    - Example - Calculate CLIP Embeddings with Roboflow Inference: examples/image_embeddings_roboflow.md
     - Example - Multi-lingual semantic search: notebooks/multi_lingual_example.ipynb
     - Example - MultiModal CLIP Embeddings: notebooks/DisappearingEmbeddingFunction.ipynb
   - 🔍 Python full-text search: fts.md

docs/src/embeddings/default_embedding_functions.md

@@ -1,7 +1,9 @@
 There are various Embedding functions available out of the box with lancedb. We're working on supporting other popular embedding APIs.
 
 ## Text Embedding Functions
-Here are the text embedding functions registered by default
+Here are the text embedding functions registered by default.
+Embedding functions have inbuilt rate limit handler wrapper for source and query embedding function calls that retry with exponential standoff. 
+Each `EmbeddingFunction` implementation automatically takes `max_retries` as an argument which has the deafult value of 7.
 
 ### Sentence Transformers
 Here are the parameters that you can set when registering a `sentence-transformers` object, and their default values:
@@ -66,6 +68,56 @@ actual = table.search(query).limit(1).to_pydantic(Words)[0]
 print(actual.text)
 ```
 
+### Instructor Embeddings
+Instructor is an instruction-finetuned text embedding model that can generate text embeddings tailored to any task (e.g., classification, retrieval, clustering, text evaluation, etc.) and domains (e.g., science, finance, etc.) by simply providing the task instruction, without any finetuning
+
+If you want to calculate customized embeddings for specific sentences, you may follow the unified template to write instructions:
+
+                          Represent the `domain` `text_type` for `task_objective`:
+
+* `domain` is optional, and it specifies the domain of the text, e.g., science, finance, medicine, etc.
+* `text_type` is required, and it specifies the encoding unit, e.g., sentence, document, paragraph, etc.
+* `task_objective` is optional, and it specifies the objective of embedding, e.g., retrieve a document, classify the sentence, etc.
+
+More information about the model can be found here - https://github.com/xlang-ai/instructor-embedding
+
+| Argument | Type | Default | Description |
+|---|---|---|---|
+| `name` | `str` | "hkunlp/instructor-base" | The name of the model to use |
+| `batch_size` | `int` | `32` | The batch size to use when generating embeddings |
+| `device` | `str` | `"cpu"` | The device to use when generating embeddings |
+| `show_progress_bar` | `bool` | `True` | Whether to show a progress bar when generating embeddings |
+| `normalize_embeddings` | `bool` | `True` | Whether to normalize the embeddings |
+| `quantize` | `bool` | `False` | Whether to quantize the model |
+| `source_instruction` | `str` | `"represent the docuement for retreival"` | The instruction for the source column |
+| `query_instruction` | `str` | `"represent the document for retreiving the most similar documents"` | The instruction for the query |
+
+
+
+```python
+import lancedb
+from lancedb.pydantic import LanceModel, Vector
+from lancedb.embeddings import get_registry, InstuctorEmbeddingFunction
+
+instructor = get_registry().get("instructor").create(
+                            source_instruction="represent the docuement for retreival",
+                            query_instruction="represent the document for retreiving the most similar documents"
+                            )
+
+class Schema(LanceModel):
+    vector: Vector(instructor.ndims()) = instructor.VectorField()
+    text: str = instructor.SourceField()
+
+db = lancedb.connect("~/.lancedb")
+tbl = db.create_table("test", schema=Schema, mode="overwrite")
+
+texts = [{"text": "Capitalism has been dominant in the Western world since the end of feudalism, but most feel[who?] that..."},
+        {"text": "The disparate impact theory is especially controversial under the Fair Housing Act because the Act..."},
+        {"text": "Disparate impact in United States labor law refers to practices in employment, housing, and other areas that.."}]
+
+tbl.add(texts)
+```
+
 ## Multi-modal embedding functions
 Multi-modal embedding functions allow you query your table using both images and text.
 

docs/src/embeddings/embedding_functions.md

@@ -57,6 +57,19 @@ query_image = Image.open(p)
 table.search(query_image)
 
 ```
+### Rate limit Handling
+`EmbeddingFunction` class wraps the calls for source and query embedding generation inside a rate limit handler that retries the requests with exponential backoff after successive failures. By default the maximum retires is set to 7. You can tune it by setting it to a different number or disable it by setting it to 0.
+Example
+----
+
+```python
+clip = registry.get("open-clip").create() # Defaults to 7 max retries
+clip = registry.get("open-clip").create(max_retries=10) # Increase max retries to 10
+clip = registry.get("open-clip").create(max_retries=0) # Retries disabled
+````
+
+NOTE:
+Embedding functions can also fail due to other errors that have nothing to do with rate limits. This is why the error is also logged.
 
 ### A little fun with PyDantic
 LanceDB is integrated with PyDantic. Infact we've used the integration in the above example to define the schema. It is also being used behing the scene by the embdding function API to ingest useful information as table metadata.

docs/src/examples/image_embeddings_roboflow.md

@@ -0,0 +1,165 @@
+# How to Load Image Embeddings into LanceDB
+
+With the rise of Large Multimodal Models (LMMs) such as [GPT-4 Vision](https://blog.roboflow.com/gpt-4-vision/), the need for storing image embeddings is growing. The most effective way to store text and image embeddings is in a vector database such as LanceDB. Vector databases are a special kind of data store that enables efficient search over stored embeddings. 
+
+[CLIP](https://blog.roboflow.com/openai-clip/), a multimodal model developed by OpenAI, is commonly used to calculate image embeddings. These embeddings can then be used with a vector database to build a semantic search engine that you can query using images or text. For example, you could use LanceDB and CLIP embeddings to build a search engine for a database of folders.
+
+In this guide, we are going to show you how to use Roboflow Inference to load image embeddings into LanceDB. Without further ado, let’s get started!
+
+## Step #1: Install Roboflow Inference
+
+[Roboflow Inference](https://inference.roboflow.com) enables you to run state-of-the-art computer vision models with minimal configuration. Inference supports a range of models, from fine-tuned object detection, classification, and segmentation models to foundation models like CLIP. We will use Inference to calculate CLIP image embeddings.
+
+Inference provides a HTTP API through which you can run vision models.
+
+Inference powers the Roboflow hosted API, and is available as an open source utility. In this guide, we are going to run Inference locally, which enables you to calculate CLIP embeddings on your own hardware. We will also show you how to use the hosted Roboflow CLIP API, which is ideal if you need to scale and do not want to manage a system for calculating embeddings.
+
+To get started, first install the Inference CLI:
+
+```
+pip install inference-cli
+```
+
+Next, install Docker. Refer to the official Docker installation instructions for your operating system to get Docker set up. Once Docker is ready, you can start Inference using the following command:
+
+```
+inference server start
+```
+
+An Inference server will start running at ‘http://localhost:9001’.
+
+## Step #2: Set Up a LanceDB Vector Database
+
+Now that we have Inference running, we can set up a LanceDB vector database. You can run LanceDB in JavaScript and Python. For this guide, we will use the Python API. But, you can take the HTTP requests we make below and change them to JavaScript if required.
+
+For this guide, we are going to search the [COCO 128 dataset](https://universe.roboflow.com/team-roboflow/coco-128), which contains a wide range of objects. The variability in objects present in this dataset makes it a good dataset to demonstrate the capabilities of vector search. If you want to use this dataset, you can download [COCO 128 from Roboflow Universe](https://universe.roboflow.com/team-roboflow/coco-128). With that said, you can search whatever folder of images you want.
+
+Once you have a dataset ready, install LanceDB with the following command:
+
+```
+pip install lancedb
+```
+
+We also need to install a specific commit of `tantivy`, a dependency of the LanceDB full text search engine we will use later in this guide:
+
+```
+pip install tantivy@git+https://github.com/quickwit-oss/tantivy-py#164adc87e1a033117001cf70e38c82a53014d985
+```
+
+Create a new Python file and add the following code:
+
+```python
+import cv2
+import supervision as sv
+import requests
+
+import lancedb
+
+db = lancedb.connect("./embeddings")
+
+IMAGE_DIR = "images/"
+API_KEY = os.environ.get("ROBOFLOW_API_KEY")
+SERVER_URL = "http://localhost:9001"
+
+results = []
+
+for i, image in enumerate(os.listdir(IMAGE_DIR)):
+    infer_clip_payload = {
+        #Images can be provided as urls or as base64 encoded strings
+        "image": {
+            "type": "base64",
+            "value": base64.b64encode(open(IMAGE_DIR + image, "rb").read()).decode("utf-8"),
+        },
+    }
+
+    res = requests.post(
+        f"{SERVER_URL}/clip/embed_image?api_key={API_KEY}",
+        json=infer_clip_payload,
+    )
+
+    embeddings = res.json()['embeddings']
+
+    print("Calculated embedding for image: ", image)
+
+    image = {"vector": embeddings[0], "name": os.path.join(IMAGE_DIR, image)}
+
+    results.append(image)
+
+tbl = db.create_table("images", data=results)
+
+tbl.create_fts_index("name")
+```
+
+To use the code above, you will need a Roboflow API key. [Learn how to retrieve a Roboflow API key](https://docs.roboflow.com/api-reference/authentication#retrieve-an-api-key). Run the following command to set up your API key in your environment:
+
+```
+export ROBOFLOW_API_KEY=""
+```
+
+Replace the `IMAGE_DIR` value with the folder in which you are storing the images for which you want to calculate embeddings. If you want to use the Roboflow CLIP API to calculate embeddings, replace the `SERVER_URL` value with `https://infer.roboflow.com`.
+
+Run the script above to create a new LanceDB database. This database will be stored on your local machine. The database will be called `embeddings` and the table will be called `images`.
+
+The script above calculates all embeddings for a folder then creates a new table. To add additional images, use the following code:
+
+```python
+def make_batches():
+    for i in range(5):
+        yield [
+                {"vector": [3.1, 4.1], "name": "image1.png"},
+                {"vector": [5.9, 26.5], "name": "image2.png"}
+            ]
+
+tbl = db.open_table("images")
+tbl.add(make_batches())
+```
+
+Replacing the `make_batches()` function with code to load embeddings for images.
+
+## Step #3: Run a Search Query
+
+We are now ready to run a search query. To run a search query, we need a text embedding that represents a text query. We can use this embedding to search our LanceDB database for an entry.
+
+Let’s calculate a text embedding for the query “cat”, then run a search query:
+
+```python
+infer_clip_payload = {
+    "text": "cat",
+}
+
+res = requests.post(
+    f"{SERVER_URL}/clip/embed_text?api_key={API_KEY}",
+    json=infer_clip_payload,
+)
+
+embeddings = res.json()['embeddings']
+
+df = tbl.search(embeddings[0]).limit(3).to_list()
+
+print("Results:")
+
+for i in df:
+    print(i["name"])
+```
+
+This code will search for the three images most closely related to the prompt “cat”. The names of the most similar three images will be printed to the console. Here are the three top results:
+
+```
+dataset/images/train/000000000650_jpg.rf.1b74ba165c5a3513a3211d4a80b69e1c.jpg
+dataset/images/train/000000000138_jpg.rf.af439ef1c55dd8a4e4b142d186b9c957.jpg
+dataset/images/train/000000000165_jpg.rf.eae14d5509bf0c9ceccddbb53a5f0c66.jpg
+```
+
+Let’s open the top image:
+
+![Cat](https://media.roboflow.com/cat_lancedb.jpg)
+
+The top image was a cat. Our search was successful.
+
+## Conclusion
+
+LanceDB is a vector database that you can use to store and efficiently search your image embeddings. You can use Roboflow Inference, a scalable computer vision inference server, to calculate CLIP embeddings that you can store in LanceDB.
+
+You can use Inference and LanceDB together to build a range of applications with image embeddings, from a media search engine to a retrieval-augmented generation pipeline for use with LMMs.
+
+To learn more about Inference and its capabilities, refer to the Inference documentation.

docs/src/robots.txt

@@ -0,0 +1 @@
+User-agent: *

node/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "vectordb",
-  "version": "0.3.7",
+  "version": "0.3.8",
   "lockfileVersion": 2,
   "requires": true,
   "packages": {
     "": {
       "name": "vectordb",
-      "version": "0.3.7",
+      "version": "0.3.8",
       "cpu": [
         "x64",
         "arm64"
@@ -53,11 +53,11 @@
         "uuid": "^9.0.0"
       },
       "optionalDependencies": {
-        "@lancedb/vectordb-darwin-arm64": "0.3.7",
-        "@lancedb/vectordb-darwin-x64": "0.3.7",
-        "@lancedb/vectordb-linux-arm64-gnu": "0.3.7",
-        "@lancedb/vectordb-linux-x64-gnu": "0.3.7",
-        "@lancedb/vectordb-win32-x64-msvc": "0.3.7"
+        "@lancedb/vectordb-darwin-arm64": "0.3.8",
+        "@lancedb/vectordb-darwin-x64": "0.3.8",
+        "@lancedb/vectordb-linux-arm64-gnu": "0.3.8",
+        "@lancedb/vectordb-linux-x64-gnu": "0.3.8",
+        "@lancedb/vectordb-win32-x64-msvc": "0.3.8"
       }
     },
     "node_modules/@apache-arrow/ts": {
@@ -316,22 +316,10 @@
         "@jridgewell/sourcemap-codec": "^1.4.10"
       }
     },
-    "node_modules/@lancedb/vectordb-darwin-arm64": {
-      "version": "0.3.7",
-      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-darwin-arm64/-/vectordb-darwin-arm64-0.3.7.tgz",
-      "integrity": "sha512-QsDxcbhrumJg+Cyflpnj8EY+bZojbco5K7VSeKvguqeXUGb62ksyOZuUTCn2sqJaCgy1KZ1qC5U8jBqfgZHc2w==",
-      "cpu": [
-        "arm64"
-      ],
-      "optional": true,
-      "os": [
-        "darwin"
-      ]
-    },
     "node_modules/@lancedb/vectordb-darwin-x64": {
-      "version": "0.3.7",
-      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-darwin-x64/-/vectordb-darwin-x64-0.3.7.tgz",
-      "integrity": "sha512-fgv10kI04UycgpmhJLUcCswgvSdgsGuj65o+W5usmVdxYZiWpoXBBXRkWYMjUX5RNe3mY1Ff6QPBbToR0WkSUA==",
+      "version": "0.3.8",
+      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-darwin-x64/-/vectordb-darwin-x64-0.3.8.tgz",
+      "integrity": "sha512-PCJwJ2oV0yTq0XJryMMjLad14i9s6xRolFZ1M4EZtgN16X/n/m0xTZjU8Y95Fj28tPFMgd4Pmgtc/TWuEBxW8A==",
       "cpu": [
         "x64"
       ],
@@ -341,9 +329,9 @@
       ]
     },
     "node_modules/@lancedb/vectordb-linux-arm64-gnu": {
-      "version": "0.3.7",
-      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-linux-arm64-gnu/-/vectordb-linux-arm64-gnu-0.3.7.tgz",
-      "integrity": "sha512-pvw+31+VKEH3YmS/GLKzEGt/Y2+c/IaE6JL6tIjXi2KY+ZcWuyyXpYnYiHHDw2EP7ubKj6+fKIG1P9tlxMcGMQ==",
+      "version": "0.3.8",
+      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-linux-arm64-gnu/-/vectordb-linux-arm64-gnu-0.3.8.tgz",
+      "integrity": "sha512-P+ZvI9+g8MDjPvz5+HPNFCCPNAvCDpCfIvKqEiTGEm2Sk5I0meIxRX4VGEnLGcQZmF1LUnVzhKV9+Rkiqd4JIQ==",
       "cpu": [
         "arm64"
       ],
@@ -353,9 +341,9 @@
       ]
     },
     "node_modules/@lancedb/vectordb-linux-x64-gnu": {
-      "version": "0.3.7",
-      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-linux-x64-gnu/-/vectordb-linux-x64-gnu-0.3.7.tgz",
-      "integrity": "sha512-kHFURhfhJRqw4k1auseqQgOzAHB4oYpyzLCX3TCR3uTxqRQ7gFxxlO0TnIcwNRqLcGb9GmWxWWoR8k1CdCXrMw==",
+      "version": "0.3.8",
+      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-linux-x64-gnu/-/vectordb-linux-x64-gnu-0.3.8.tgz",
+      "integrity": "sha512-E7opS6JuNpyvej0JZ+DJxplnnFp543dlPW0hNxoxsndflo9NeeAa1AIsNQSCIABWlfsQbGxXPYrvsOKHbzAIdw==",
       "cpu": [
         "x64"
       ],
@@ -365,9 +353,9 @@
       ]
     },
     "node_modules/@lancedb/vectordb-win32-x64-msvc": {
-      "version": "0.3.7",
-      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-win32-x64-msvc/-/vectordb-win32-x64-msvc-0.3.7.tgz",
-      "integrity": "sha512-zWfZ557v2Y+93dVrmqqnbiLeTOb0ptunAG0zGjyE+3oyi8j/4+bL56Fdv94k+dfNF4KrcqcULEcZhKik3/FQ9w==",
+      "version": "0.3.8",
+      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-win32-x64-msvc/-/vectordb-win32-x64-msvc-0.3.8.tgz",
+      "integrity": "sha512-RAL8U46UE12ksO3VAnnLlfxDd4wxZpJNFYtXjkacKL4ud9PkAJC4FBJpD7EFP9c7LEY3IlJPtvAp5Ax9LGWFeA==",
       "cpu": [
         "x64"
       ],
@@ -4868,34 +4856,28 @@
         "@jridgewell/sourcemap-codec": "^1.4.10"
       }
     },
-    "@lancedb/vectordb-darwin-arm64": {
-      "version": "0.3.7",
-      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-darwin-arm64/-/vectordb-darwin-arm64-0.3.7.tgz",
-      "integrity": "sha512-QsDxcbhrumJg+Cyflpnj8EY+bZojbco5K7VSeKvguqeXUGb62ksyOZuUTCn2sqJaCgy1KZ1qC5U8jBqfgZHc2w==",
-      "optional": true
-    },
     "@lancedb/vectordb-darwin-x64": {
-      "version": "0.3.7",
-      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-darwin-x64/-/vectordb-darwin-x64-0.3.7.tgz",
-      "integrity": "sha512-fgv10kI04UycgpmhJLUcCswgvSdgsGuj65o+W5usmVdxYZiWpoXBBXRkWYMjUX5RNe3mY1Ff6QPBbToR0WkSUA==",
+      "version": "0.3.8",
+      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-darwin-x64/-/vectordb-darwin-x64-0.3.8.tgz",
+      "integrity": "sha512-PCJwJ2oV0yTq0XJryMMjLad14i9s6xRolFZ1M4EZtgN16X/n/m0xTZjU8Y95Fj28tPFMgd4Pmgtc/TWuEBxW8A==",
       "optional": true
     },
     "@lancedb/vectordb-linux-arm64-gnu": {
-      "version": "0.3.7",
-      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-linux-arm64-gnu/-/vectordb-linux-arm64-gnu-0.3.7.tgz",
-      "integrity": "sha512-pvw+31+VKEH3YmS/GLKzEGt/Y2+c/IaE6JL6tIjXi2KY+ZcWuyyXpYnYiHHDw2EP7ubKj6+fKIG1P9tlxMcGMQ==",
+      "version": "0.3.8",
+      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-linux-arm64-gnu/-/vectordb-linux-arm64-gnu-0.3.8.tgz",
+      "integrity": "sha512-P+ZvI9+g8MDjPvz5+HPNFCCPNAvCDpCfIvKqEiTGEm2Sk5I0meIxRX4VGEnLGcQZmF1LUnVzhKV9+Rkiqd4JIQ==",
       "optional": true
     },
     "@lancedb/vectordb-linux-x64-gnu": {
-      "version": "0.3.7",
-      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-linux-x64-gnu/-/vectordb-linux-x64-gnu-0.3.7.tgz",
-      "integrity": "sha512-kHFURhfhJRqw4k1auseqQgOzAHB4oYpyzLCX3TCR3uTxqRQ7gFxxlO0TnIcwNRqLcGb9GmWxWWoR8k1CdCXrMw==",
+      "version": "0.3.8",
+      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-linux-x64-gnu/-/vectordb-linux-x64-gnu-0.3.8.tgz",
+      "integrity": "sha512-E7opS6JuNpyvej0JZ+DJxplnnFp543dlPW0hNxoxsndflo9NeeAa1AIsNQSCIABWlfsQbGxXPYrvsOKHbzAIdw==",
       "optional": true
     },
     "@lancedb/vectordb-win32-x64-msvc": {
-      "version": "0.3.7",
-      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-win32-x64-msvc/-/vectordb-win32-x64-msvc-0.3.7.tgz",
-      "integrity": "sha512-zWfZ557v2Y+93dVrmqqnbiLeTOb0ptunAG0zGjyE+3oyi8j/4+bL56Fdv94k+dfNF4KrcqcULEcZhKik3/FQ9w==",
+      "version": "0.3.8",
+      "resolved": "https://registry.npmjs.org/@lancedb/vectordb-win32-x64-msvc/-/vectordb-win32-x64-msvc-0.3.8.tgz",
+      "integrity": "sha512-RAL8U46UE12ksO3VAnnLlfxDd4wxZpJNFYtXjkacKL4ud9PkAJC4FBJpD7EFP9c7LEY3IlJPtvAp5Ax9LGWFeA==",
       "optional": true
     },
     "@neon-rs/cli": {

node/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vectordb",
-  "version": "0.3.7",
+  "version": "0.3.9",
   "description": " Serverless, low-latency vector database for AI applications",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -81,10 +81,10 @@
     }
   },
   "optionalDependencies": {
-    "@lancedb/vectordb-darwin-arm64": "0.3.7",
-    "@lancedb/vectordb-darwin-x64": "0.3.7",
-    "@lancedb/vectordb-linux-arm64-gnu": "0.3.7",
-    "@lancedb/vectordb-linux-x64-gnu": "0.3.7",
-    "@lancedb/vectordb-win32-x64-msvc": "0.3.7"
+    "@lancedb/vectordb-darwin-arm64": "0.3.9",
+    "@lancedb/vectordb-darwin-x64": "0.3.9",
+    "@lancedb/vectordb-linux-arm64-gnu": "0.3.9",
+    "@lancedb/vectordb-linux-x64-gnu": "0.3.9",
+    "@lancedb/vectordb-win32-x64-msvc": "0.3.9"
   }
 }

node/src/query.ts

@@ -32,6 +32,7 @@ export class Query<T = number[]> {
   private _select?: string[]
   private _filter?: string
   private _metricType?: MetricType
+  private _prefilter: boolean
   protected readonly _embeddings?: EmbeddingFunction<T>
 
   constructor (query: T, tbl?: any, embeddings?: EmbeddingFunction<T>) {
@@ -44,6 +45,7 @@ export class Query<T = number[]> {
     this._filter = undefined
     this._metricType = undefined
     this._embeddings = embeddings
+    this._prefilter = false
   }
 
   /***
@@ -102,6 +104,11 @@ export class Query<T = number[]> {
     return this
   }
 
+  prefilter (value: boolean): Query<T> {
+    this._prefilter = value
+    return this
+  }
+
   /**
      * Execute the query and return the results as an Array of Objects
      */

node/src/remote/client.ts

@@ -38,6 +38,7 @@ export class HttpLancedbClient {
     vector: number[],
     k: number,
     nprobes: number,
+    prefilter: boolean,
     refineFactor?: number,
     columns?: string[],
     filter?: string
@@ -50,7 +51,8 @@ export class HttpLancedbClient {
                 nprobes,
                 refineFactor,
                 columns,
-                filter
+                filter,
+                prefilter
               },
               {
                 headers: {

node/src/remote/index.ts

@@ -154,6 +154,7 @@ export class RemoteQuery<T = number[]> extends Query<T> {
       queryVector,
       (this as any)._limit,
       (this as any)._nprobes,
+      (this as any)._prefilter,
       (this as any)._refineFactor,
       (this as any)._select,
       (this as any)._filter

node/src/test/test.ts

@@ -102,6 +102,20 @@ describe('LanceDB client', function () {
       assertResults(results)
     })
 
+    it('should correctly process prefilter/postfilter', async function () {
+      const uri = await createTestDB(16, 300)
+      const con = await lancedb.connect(uri)
+      const table = await con.openTable('vectors')
+      await table.createIndex({ type: 'ivf_pq', column: 'vector', num_partitions: 2, max_iters: 2, num_sub_vectors: 2 })
+      // post filter should return less than the limit
+      let results = await table.search(new Array(16).fill(0.1)).limit(10).filter('id >= 10').prefilter(false).execute()
+      assert.isTrue(results.length < 10)
+
+      // pre filter should return exactly the limit
+      results = await table.search(new Array(16).fill(0.1)).limit(10).filter('id >= 10').prefilter(true).execute()
+      assert.isTrue(results.length === 10)
+    })
+
     it('select only a subset of columns', async function () {
       const uri = await createTestDB()
       const con = await lancedb.connect(uri)
@@ -282,7 +296,8 @@ describe('LanceDB client', function () {
       )
       const table = await con.createTable({ name: 'vectors', schema })
       await table.add([{ vector: Array(128).fill(0.1) }])
-      await table.delete('vector IS NOT NULL')
+      // https://github.com/lancedb/lance/issues/1635
+      await table.delete('true')
       const result = await table.search(Array(128).fill(0.1)).execute()
       assert.isEmpty(result)
     })

python/.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.3.3
+current_version = 0.3.4
 commit = True
 message = [python] Bump version: {current_version} → {new_version}
 tag = True

python/lancedb/remote/db.py

@@ -28,6 +28,7 @@ from ..pydantic import LanceModel
 from ..table import Table, _sanitize_data
 from .arrow import to_ipc_binary
 from .client import ARROW_STREAM_CONTENT_TYPE, RestfulLanceDBClient
+from .errors import LanceDBClientError
 
 
 class RemoteDBConnection(DBConnection):
@@ -101,11 +102,12 @@ class RemoteDBConnection(DBConnection):
             self._loop.run_until_complete(
                 self._client.post(f"/v1/table/{name}/describe/")
             )
-        except Exception:
-            logging.error(
-                "Table {name} does not exist."
-                "Please first call db.create_table({name}, data)"
-            )
+        except LanceDBClientError as err:
+            if str(err).startswith("Not found"):
+                logging.error(
+                    f"Table {name} does not exist. "
+                    f"Please first call db.create_table({name}, data)"
+                )
         return RemoteTable(self, name)
 
     @override

python/lancedb/remote/table.py

@@ -71,8 +71,62 @@ class RemoteTable(Table):
         vector_column_name: str = VECTOR_COLUMN_NAME,
         replace: bool = True,
         accelerator: Optional[str] = None,
+        index_cache_size: Optional[int] = None,
     ):
-        raise NotImplementedError
+        """Create an index on the table.
+        Currently, the only parameters that matter are
+        the metric and the vector column name.
+
+        Parameters
+        ----------
+        metric : str
+            The metric to use for the index. Default is "L2".
+        num_partitions : int
+            The number of partitions to use for the index. Default is 256.
+        num_sub_vectors : int
+            The number of sub-vectors to use for the index. Default is 96.
+        vector_column_name : str
+            The name of the vector column. Default is "vector".
+        replace : bool
+            Whether to replace the existing index. Default is True.
+        accelerator : str, optional
+            If set, use the given accelerator to create the index.
+            Default is None. Currently not supported.
+        index_cache_size : int, optional
+            The size of the index cache in number of entries. Default value is 256.
+
+        Examples
+        --------
+        import lancedb
+        import uuid
+        from lancedb.schema import vector
+        conn = lancedb.connect("db://...", api_key="...", region="...")
+        table_name = uuid.uuid4().hex
+        schema = pa.schema(
+            [
+                    pa.field("id", pa.uint32(), False),
+                    pa.field("vector", vector(128), False),
+                    pa.field("s", pa.string(), False),
+            ]
+        )
+        table = conn.create_table(
+            table_name,
+            schema=schema,
+        )
+        table.create_index()
+        """
+        index_type = "vector"
+
+        data = {
+            "column": vector_column_name,
+            "index_type": index_type,
+            "metric_type": metric,
+            "index_cache_size": index_cache_size,
+        }
+        resp = self._conn._loop.run_until_complete(
+            self._conn._client.post(f"/v1/table/{self._name}/create_index/", data=data)
+        )
+        return resp
 
     def add(
         self,

python/lancedb/table.py

@@ -188,6 +188,7 @@ class Table(ABC):
         vector_column_name: str = VECTOR_COLUMN_NAME,
         replace: bool = True,
         accelerator: Optional[str] = None,
+        index_cache_size: Optional[int] = None,
     ):
         """Create an index on the table.
 
@@ -212,6 +213,8 @@ class Table(ABC):
         accelerator: str, default None
             If set, use the given accelerator to create the index.
             Only support "cuda" for now.
+        index_cache_size : int, optional
+            The size of the index cache in number of entries. Default value is 256.
         """
         raise NotImplementedError
 
@@ -556,6 +559,7 @@ class LanceTable(Table):
         vector_column_name=VECTOR_COLUMN_NAME,
         replace: bool = True,
         accelerator: Optional[str] = None,
+        index_cache_size: Optional[int] = None,
     ):
         """Create an index on the table."""
         self._dataset.create_index(
@@ -566,6 +570,7 @@ class LanceTable(Table):
             num_sub_vectors=num_sub_vectors,
             replace=replace,
             accelerator=accelerator,
+            index_cache_size=index_cache_size,
         )
         self._reset_dataset()
         register_event("create_index")

python/pyproject.toml

@@ -1,9 +1,9 @@
 [project]
 name = "lancedb"
-version = "0.3.3"
+version = "0.3.4"
 dependencies = [
     "deprecation",
-    "pylance==0.8.10",
+    "pylance==0.8.17",
     "ratelimiter~=1.0",
     "retry>=0.9.2",
     "tqdm>=4.1.0",

python/tests/test_db.py

@@ -301,6 +301,7 @@ def test_replace_index(tmp_path):
         num_partitions=2,
         num_sub_vectors=4,
         replace=True,
+        index_cache_size=10,
     )
 
 

python/tests/test_remote_db.py

@@ -26,6 +26,9 @@ class FakeLanceDBClient:
         t = pa.schema([]).empty_table()
         return VectorQueryResult(t)
 
+    async def post(self, path: str):
+        pass
+
 
 def test_remote_db():
     conn = lancedb.connect("db://client-will-be-injected", api_key="fake")

python/tests/test_table.py

@@ -213,6 +213,7 @@ def test_create_index_method():
                 num_sub_vectors=96,
                 vector_column_name="vector",
                 replace=True,
+                index_cache_size=256,
             )
 
             # Check that the _dataset.create_index method was called
@@ -225,6 +226,7 @@ def test_create_index_method():
                 num_sub_vectors=96,
                 replace=True,
                 accelerator=None,
+                index_cache_size=256,
             )
 
 

rust/ffi/node/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "vectordb-node"
-version = "0.3.7"
+version = "0.3.9"
 description = "Serverless, low-latency vector database for AI applications"
 license = "Apache-2.0"
 edition = "2018"
@@ -19,6 +19,7 @@ once_cell = "1"
 futures = "0.3"
 half = { workspace = true }
 lance = { workspace = true }
+lance-index = { workspace = true }
 lance-linalg = { workspace = true }
 vectordb = { path = "../../vectordb" }
 tokio = { version = "1.23", features = ["rt-multi-thread"] }

rust/ffi/node/src/index/vector.rs

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use lance::index::vector::{ivf::IvfBuildParams, pq::PQBuildParams};
+use lance_index::vector::{ivf::IvfBuildParams, pq::PQBuildParams};
 use lance_linalg::distance::MetricType;
 use neon::context::FunctionContext;
 use neon::prelude::*;

rust/ffi/node/src/query.rs

@@ -48,6 +48,8 @@ impl JsQuery {
             .map(|s| s.value(&mut cx))
             .map(|s| MetricType::try_from(s.as_str()).unwrap());
 
+        let prefilter = query_obj.get::<JsBoolean, _, _>(&mut cx, "_prefilter")?.value(&mut cx);
+
         let is_electron = cx
             .argument::<JsBoolean>(1)
             .or_throw(&mut cx)?
@@ -69,7 +71,8 @@ impl JsQuery {
                 .nprobes(nprobes)
                 .filter(filter)
                 .metric_type(metric_type)
-                .select(select);
+                .select(select)
+                .prefilter(prefilter);
             let record_batch_stream = builder.execute();
             let results = record_batch_stream
                 .and_then(|stream| {

rust/vectordb/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "vectordb"
-version = "0.3.7"
+version = "0.3.9"
 edition = "2021"
 description = "LanceDB: A serverless, low-latency vector database for AI applications"
 license = "Apache-2.0"
@@ -21,6 +21,7 @@ object_store = { workspace = true }
 snafu = { workspace = true }
 half = { workspace = true }
 lance = { workspace = true }
+lance-index = { workspace = true }
 lance-linalg = { workspace = true }
 lance-testing = { workspace = true }
 tokio = { version = "1.23", features = ["rt-multi-thread"] }

rust/vectordb/src/index/vector.rs

@@ -13,9 +13,9 @@
 // limitations under the License.
 
 use lance::format::{Index, Manifest};
-use lance::index::vector::ivf::IvfBuildParams;
 use lance::index::vector::pq::PQBuildParams;
 use lance::index::vector::VectorIndexParams;
+use lance_index::vector::ivf::IvfBuildParams;
 use lance_linalg::distance::MetricType;
 
 pub trait VectorIndexBuilder {
@@ -136,9 +136,9 @@ impl VectorIndex {
 mod tests {
     use super::*;
 
-    use lance::index::vector::ivf::IvfBuildParams;
-    use lance::index::vector::pq::PQBuildParams;
     use lance::index::vector::StageParams;
+    use lance_index::vector::ivf::IvfBuildParams;
+    use lance_index::vector::pq::PQBuildParams;
 
     use crate::index::vector::{IvfPQIndexBuilder, VectorIndexBuilder};
 

rust/vectordb/src/query.rs

@@ -32,6 +32,7 @@ pub struct Query {
     pub refine_factor: Option<u32>,
     pub metric_type: Option<MetricType>,
     pub use_index: bool,
+    pub prefilter: bool,
 }
 
 impl Query {
@@ -56,6 +57,7 @@ impl Query {
             use_index: true,
             filter: None,
             select: None,
+            prefilter: false,
         }
     }
 
@@ -74,6 +76,8 @@ impl Query {
         )?;
         scanner.nprobs(self.nprobes);
         scanner.use_index(self.use_index);
+        scanner.prefilter(self.prefilter);
+
         self.select.as_ref().map(|p| scanner.project(p.as_slice()));
         self.filter.as_ref().map(|f| scanner.filter(f));
         self.refine_factor.map(|rf| scanner.refine(rf));
@@ -158,6 +162,11 @@ impl Query {
         self.select = columns;
         self
     }
+
+    pub fn prefilter(mut self, prefilter: bool) -> Query {
+        self.prefilter = prefilter;
+        self
+    }
 }
 
 #[cfg(test)]
@@ -167,7 +176,9 @@ mod tests {
     use super::*;
     use arrow_array::{Float32Array, RecordBatch, RecordBatchIterator, RecordBatchReader};
     use arrow_schema::{DataType, Field as ArrowField, Schema as ArrowSchema};
+    use futures::StreamExt;
     use lance::dataset::Dataset;
+    use lance_testing::datagen::{BatchGenerator, IncrementingInt32, RandomVector};
 
     use crate::query::Query;
 
@@ -200,13 +211,43 @@ mod tests {
 
     #[tokio::test]
     async fn test_execute() {
-        let batches = make_test_batches();
-        let ds = Dataset::write(batches, "memory://foo", None).await.unwrap();
+        let batches = make_non_empty_batches();
+        let ds = Arc::new(Dataset::write(batches, "memory://foo", None).await.unwrap());
 
-        let vector = Float32Array::from_iter_values([0.1; 128]);
-        let query = Query::new(Arc::new(ds), vector.clone());
-        let result = query.execute().await;
-        assert_eq!(result.is_ok(), true);
+        let vector = Float32Array::from_iter_values([0.1; 4]);
+
+        let query = Query::new(ds.clone(), vector.clone());
+        let result = query
+            .limit(10)
+            .filter(Some("id % 2 == 0".to_string()))
+            .execute()
+            .await;
+        let mut stream = result.expect("should have result");
+        // should only have one batch
+        while let Some(batch) = stream.next().await {
+            // post filter should have removed some rows
+            assert!(batch.expect("should be Ok").num_rows() < 10);
+        }
+
+        let query = Query::new(ds, vector.clone());
+        let result = query
+            .limit(10)
+            .filter(Some("id % 2 == 0".to_string()))
+            .prefilter(true)
+            .execute()
+            .await;
+        let mut stream = result.expect("should have result");
+        // should only have one batch
+        while let Some(batch) = stream.next().await {
+            // pre filter should return 10 rows
+            assert!(batch.expect("should be Ok").num_rows() == 10);
+        }
+    }
+
+    fn make_non_empty_batches() -> impl RecordBatchReader + Send + 'static {
+        let vec = Box::new(RandomVector::new().named("vector".to_string()));
+        let id = Box::new(IncrementingInt32::new().named("id".to_string()));
+        BatchGenerator::new().col(vec).col(id).batch(512)
     }
 
     fn make_test_batches() -> impl RecordBatchReader + Send + 'static {

rust/vectordb/src/table.rs

@@ -13,6 +13,8 @@
 // limitations under the License.
 
 use chrono::Duration;
+use lance::dataset::builder::DatasetBuilder;
+use lance_index::IndexType;
 use std::sync::Arc;
 
 use arrow_array::{Float32Array, RecordBatchReader};
@@ -22,7 +24,7 @@ use lance::dataset::optimize::{
     compact_files, CompactionMetrics, CompactionOptions, IndexRemapperOptions,
 };
 use lance::dataset::{Dataset, WriteParams};
-use lance::index::{DatasetIndexExt, IndexType};
+use lance::index::DatasetIndexExt;
 use lance::io::object_store::WrappingObjectStore;
 use std::path::Path;
 
@@ -96,7 +98,10 @@ impl Table {
             Some(wrapper) => params.patch_with_store_wrapper(wrapper)?,
             None => params,
         };
-        let dataset = Dataset::open_with_params(uri, &params)
+
+        let dataset = DatasetBuilder::from_uri(uri)
+            .with_read_params(params)
+            .load()
             .await
             .map_err(|e| match e {
                 lance::Error::DatasetNotFound { .. } => Error::TableNotFound {
@@ -414,9 +419,9 @@ mod tests {
     use arrow_data::ArrayDataBuilder;
     use arrow_schema::{DataType, Field, Schema};
     use lance::dataset::{Dataset, WriteMode};
-    use lance::index::vector::ivf::IvfBuildParams;
     use lance::index::vector::pq::PQBuildParams;
     use lance::io::object_store::{ObjectStoreParams, WrappingObjectStore};
+    use lance_index::vector::ivf::IvfBuildParams;
     use rand::Rng;
     use tempfile::tempdir;