vsag

Brand: Alibaba
Category: SDKs & Libraries
Type: Open‑source vector indexing library

Overview

vsag is an open-source C++ vector indexing library for high-dimensional similarity search, designed to handle vector datasets that may not fit entirely in memory. It implements efficient vector search algorithms, including approximate nearest neighbor (ANN) search, and provides a Python wrapper package, pyvsag, for easier integration in Python environments.

Features

• Vector similarity search

  • Supports high-dimensional similarity search and approximate nearest neighbor (ANN) queries.
  • Designed to work with vector sets of various sizes, including those that do not fit into memory.

• Indexing algorithms

  • Implements the VSAG algorithm for efficient vector indexing and search.
  • Includes SINDI algorithm optimized for sparse vector search.
  • Provides example implementations such as HNSW-based indexing (e.g., 101_index_hnsw.cpp, example_hnsw.py).

• Performance characteristics

  • SINDI algorithm for sparse vector search targets state-of-the-art performance, significantly improving QPS (queries per second) compared with prior solutions in internal tests.
  • Demonstrated efficiency on:
    • Sparse-full-inner-product tasks, with substantial QPS gains over previous SOTA and baseline algorithms on benchmark datasets.
    • gist-960-euclidean benchmarks, tested in the ann-benchmarks framework.
  • Benchmarks reported on multi-core CPU environments (e.g., Intel Xeon CPUs, AWS r6i.16xlarge), suitable for large-scale, CPU-based deployments.

• Automatic parameter generation

  • Provides methods to generate algorithm parameters based on vector dimensions and data scale.
  • Aims to let developers use the library effectively without needing deep knowledge of the underlying algorithms.

• Language support

  • C++ core implementation.
  • Python wrapper package: pyvsag available on PyPI.

• Integration & build

  • CMake integration via FetchContent for easy inclusion in C++ projects.
  • Public C++ headers available via include directory.
  • Example CMake configuration for fetching and linking vsag in user projects.
  • Build-from-source instructions provided in a DEVELOPMENT.md guide in the repository.

• Examples and usage

  • C++ and Python example programs provided in the examples directory.
  • Starter examples include:
    • 101_index_hnsw.cpp (C++)
    • example_hnsw.py (Python)

• Ecosystem adoption

  • Used by multiple database and data systems, including (as listed in the repo):
    • OceanBase
    • TuGraph
    • GreptimeDB
    • Hologres
    • PolarDB

Tech Stack

• Core language: C++ (C++11 standard or later)
• Build system: CMake (3.11+ recommended)
• Python wrapper: pyvsag (via PyPI)

Pricing

vsag is an open-source library. No pricing plans are listed; usage is free under the terms of its open-source license (see the GitHub repository for license details).

Links

• Source code & documentation: https://github.com/alipay/vsag
• Python package (pyvsag): https://pypi.org/project/pyvsag/
• Benchmarks information: https://ann-benchmarks.com/ (external benchmark framework referenced by the project)