pyglass is a library for fast inference of graph index for approximate similarity search.
- Supports multiple graph algorithms, like HNSW and NSG.
- Supports multiple hardware platforms, like X86 and ARM. Support for GPU is on the way
- No third-party library dependencies, does not rely on OpenBLAS / MKL or any other computing framework.
- Sophisticated memory management and data structure design, very low memory footprint.
- It's high performant.
sudo apt-get update && sudo apt-get install -y build-essential git python3 python3-distutils python3-venvpip3 install numpy
pip3 install pybind11bash build.shA runnable demo is at examples/demo.ipynb. It's highly recommended to try it.
Import library
>>> import glassLoad Data
>>> import numpy as np
>>> n, d = 10000, 128
>>> X = np.random.randn(n, d)
>>> Y = np.random.randn(d)Create Index pyglass supports HNSW and NSG index currently, with different quantization support
>>> index = glass.Index(index_type="HNSW", metric="L2", R=32, L=50)
>>> index = glass.Index(index_type="NSG", metric="L2", R=32, L=50, quant="SQ8U")Build Graph
>>> graph = index.build(X)Create Searcher
>>> searcher = glass.Searcher(graph=graph, data=X, metric="L2", quantizer="SQ4U")
>>> searcher.set_ef(32)(Optional) Optimize Searcher
>>> searcher.optimize()Searching
>>> ret = searcher.search(query=Y, k=10)
>>> print(ret)- FP8_E5M2
- PQ8
- SQ8
- SQ8U
- SQ6
- SQ4
- SQ4U
- SQ4UA
- SQ2U
- BinaryQuant
Rule of Thumb: Use SQ8U for indexing, and SQ4U for searching is almost always a good choice.
Glass is among one of the top performant ann algorithms on ann-benchmarks
- Change configuration file
examples/config.json - Run benchmark
python3 examples/main.py
You can cite the PyGlass repo as follows:
@misc{PyGlass,
author = {Zihao Wang},
title = {Graph Library for Approximate Similarity Search},
url = {https://github.com/zilliztech/pyglass},
year = {2025},
month = {4},
}