Scalable Vector Search

Scalable Vector Search (SVS) is a performance library for vector similarity search. Thanks to the use of Locally-adaptive Vector Quantization (LVQ) [ABHT23] and its highly optimized indexing and search algorithms, SVS provides vector similarity search:

• on billions of high-dimensional vectors,
• at high accuracy
• and state-of-the-art speed,
• while enabling the use of less memory than its alternatives.

This enables application and framework developers using similarity search to unleash its performance on Intel(R) Xeon(R) CPUs (2nd generation and newer).

SVS offers a fully-featured and yet simple Python API, compatible with most standard libraries. SVS is written in C++ to facilitate its integration into performance-critical applications.

Please note that this repository only contains the open-source portion of the SVS library, which supports all functionalities and features described in the documentation, except for our proprietary vector compression techniques, specifically LVQ [ABHT23] and Leanvec [TBAH24]. These techniques are closed-source and supported exclusively on Intel hardware. We provide shared library and PyPI package to enable these vector compression techniques in C++ and Python, respectively. For guidance on using the shared library, please refer to this example.

Performance

SVS provides state-of-the-art performance and accuracy [ABHT23] for billion-scale similarity search on standard benchmarks.

For example, for the standard billion-scale Deep-1B dataset, different configurations of SVS yield significantly increased performance (measured in queries per second, QPS) with a smaller memory footprint (horizontal axis) than the alternatives¹:

SVS is primarily optimized for large-scale similarity search but it still offers state-of-the-art performance at million-scale.

Best performance is obtained with Intel(R) Xeon(R) 6 processors (Granite Rapids), by making use of Intel(R) AVX-512 instructions, with excellent results also with 2nd through 5th gen Intel(R) Xeon(R) processors (Cascade Lake, Ice Lake, Sapphire Rapids, and Emerald Rapids).

Performance will be degraded if Intel(R) AVX-512 instructions are not available. A warning message will appear when loading the SVS Python module if the system does not support Intel(R) AVX-512 instructions.

Key Features

SVS supports:

• Similarity functions: Euclidean distance, inner product, cosine similarity.
• Vectors with individual values encoded as: float32, float16, uint8, int8.
• Vector compression (including Locally-adaptive Vector Quantization [ABHT23])
• Optimizations for Intel(R) Xeon(R) processors:
  • 2nd generation (Cascade Lake)
  • 3rd generation (Ice Lake)
  • 4th generation (Sapphire Rapids)
  • 5th generation (Emerald Rapids)
  • 6th generation (Granite Rapids)

See Roadmap for upcoming features.

Documentation

SVS documentation includes getting started tutorials with installation instructions for Python and C++ and step-by-step search examples, an API reference, as well as several guides and benchmark comparisons.

References

Reference to cite when you use SVS in a research paper:

@article{aguerrebere2023similarity,
        title={Similarity search in the blink of an eye with compressed indices},
        volume = {16},
        number = {11},
        pages = {3433--3446},
        journal = {Proceedings of the VLDB Endowment},
        author={Cecilia Aguerrebere and Ishwar Bhati and Mark Hildebrand and Mariano Tepper and Ted Willke},
        year = {2023}
}

[ABHT23] Aguerrebere, C.; Bhati I.; Hildebrand M.; Tepper M.; Willke T.:Similarity search in the blink of an eye with compressed indices. In: Proceedings of the VLDB Endowment, 16, 11, 3433 - 3446. (2023)

[TBAH24] Tepper M.; Bhati I.; Aguerrebere, C.; Hildebrand M.; Willke T.:LeanVec: Searching vectors faster by making them fit. In: Transactions on Machine Learning Research(TMLR), ISSN, 2835 - 8856. (2024)

Legal

Refer to the LICENSE file for details.