GitHub - MinishLab/vicinity: Lightweight Nearest Neighbors with Flexible Backends (original) (raw)

Vicinity is a light-weight, low-dependency vector store. It provides a simple and intuitive interface for nearest neighbor search, with support for different backends and evaluation.

There are many nearest neighbors packages and methods out there. However, we found it difficult to compare them. Every package has its own interface, quirks, and limitations, and learning a new package can be time-consuming. In addition to that, how do you effectively evaluate different packages? How do you know which one is the best for your use case?

This is where Vicinity comes in. Instead of learning a new interface for each new package or backend, Vicinity provides a unified interface for all backends. This allows you to easily experiment with different indexing methods and distance metrics and choose the best one for your use case. Vicinity also provides a simple way to evaluate the performance of different backends, allowing you to measure the queries per second and recall.

Quickstart

Install the package with:

Optionally, install specific backends and integrations, or simply install all of them with:

pip install vicinity[all]

The following code snippet demonstrates how to use Vicinity for nearest neighbor search:

import numpy as np from vicinity import Vicinity, Backend, Metric

Create some dummy data as strings or other serializable objects

items = ["triforce", "master sword", "hylian shield", "boomerang", "hookshot"] vectors = np.random.rand(len(items), 128)

Initialize the Vicinity instance (using the basic backend and cosine metric)

vicinity = Vicinity.from_vectors_and_items( vectors=vectors, items=items, backend_type=Backend.BASIC, metric=Metric.COSINE )

Create a query vector

query_vector = np.random.rand(128)

Query for nearest neighbors with a top-k search

results = vicinity.query(query_vector, k=3)

Query for nearest neighbors with a threshold search

results = vicinity.query_threshold(query_vector, threshold=0.9)

Query with a list of query vectors

query_vectors = np.random.rand(5, 128) results = vicinity.query(query_vectors, k=3)

Saving and loading a vector store:

vicinity.save('my_vector_store') vicinity = Vicinity.load('my_vector_store')

Pushing and loading a vector store from the Hugging Face Hub (note that you can optionally add the model used for generating embeddings to the metadata, e.g. vicinity.metadata["model"] = "minishlab/potion-base-8M"):

vicinity.push_to_hub(repo_id='minishlab/my-vicinity-repo') vicinity = Vicinity.load_from_hub(repo_id='minishlab/my-vicinity-repo')

Evaluating a backend:

Use the first 1000 vectors as query vectors

query_vectors = vectors[:1000]

Evaluate the Vicinity instance by measuring the queries per second and recall

qps, recall = vicinity.evaluate( full_vectors=vectors, query_vectors=query_vectors, )

Main Features

Vicinity provides the following features:

• Lightweight: Minimal dependencies and fast performance.
• Flexible Backend Support: Use different backends for vector storage and search.
• Serialization: Save and load vector stores for persistence.
• HuggingFace Hub Integration: Push and load vector stores directly to and from the HuggingFace Hub.
• Evaluation: Easily evaluate the performance of different backends.
• Easy to Use: Simple and intuitive API.

Supported Backends

The following backends are supported:

• BASIC: A simple (exact matching) flat index for vector storage and search.
• HNSW: Hierarchical Navigable Small World Graph (HNSW) for ANN search using hnswlib.
• USEARCH: ANN search using Usearch. This uses a highly optimized version of the HNSW algorithm.
• ANNOY: "Approximate Nearest Neighbors Oh Yeah" for approximate nearest neighbor search.
• PYNNDESCENT: ANN search using PyNNDescent.
• FAISS: All FAISS indexes are supported:
  • flat: Exact search.
  • ivf: Inverted file search.
  • hnsw: Hierarchical Navigable Small World Graph.
  • lsh: Locality Sensitive Hashing.
  • scalar: Scalar quantizer.
  • pq: Product Quantizer.
  • ivf_scalar: Inverted file search with scalar quantizer.
  • ivfpq: Inverted file search with product quantizer.
  • ivfpqr: Inverted file search with product quantizer and refinement.
• VOYAGER: Voyager is a library for performing fast approximate nearest-neighbor searches on an in-memory collection of vectors.

NOTE: the ANN backends do not support dynamic deletion. To delete items, you need to recreate the index. Insertion is supported in the following backends: FAISS, HNSW, and Usearch. The BASIC backend supports both insertion and deletion.

Backend Parameters

Installation

The following installation options are available:

Install the base package

pip install vicinity

Install all integrations and backends

pip install vicinity[all]

Install all integrations

pip install vicinity[integrations]

Install specific integrations

pip install vicinity[huggingface]

Install all backends

pip install vicinity[backends]

Install specific backends

pip install vicinity[annoy] pip install vicinity[faiss] pip install vicinity[hnsw] pip install vicinity[pynndescent] pip install vicinity[usearch] pip install vicinity[voyager]

License

MIT

Citing

If you use Vicinity in your research, please cite the following:

@software{minishlab2024vicinity, author = {Stephan Tulkens and {van Dongen}, Thomas}, title = {Vicinity: Lightweight Nearest Neighbors with Flexible Backends}, year = {2024}, publisher = {Zenodo}, doi = {10.5281/zenodo.17265874}, url = {https://github.com/MinishLab/vicinity}, license = {MIT} }