RaBitQ Library

Compact vectors. Accurate distances. Fast ANN search.

A research-backed C++17 library with Python bindings for 1-bit and multi-bit

vector quantization, IVF, HNSW, and SymphonyQG.

Documentation · Python package · Paper · Releases

Install

pip install rabitqlib

Prebuilt wheels support Linux x86-64 and CPython 3.11–3.14. AVX2 + FMA is the portable CPU baseline; supported AVX-512 kernels are selected at runtime.

RaBitQ across the vector-search ecosystem

Milvus	Faiss	NVIDIA cuVS	Microsoft DiskANN	VSAG
VectorChord	Volcengine OpenSearch	CockroachDB	Elasticsearch	Apache Lucene
turbopuffer	Zvec	LanceDB	Databricks	ClickHouse
Qdrant	Weaviate

Accuracy at a glance

Average and maximum relative estimation error across six datasets; lower is better. Results from the SIGMOD camera-ready paper.

Why RaBitQ?

The library supports Euclidean distance and inner product. Cosine search is available by normalizing vectors before using inner product.

RaBitQ is developed by the VectorDB group at Nanyang Technological University, Singapore. A GPU implementation is also available in cuvs_rabitq.

Python quick start

The following complete example builds a small IVF index and searches it. It uses deterministic synthetic data, so no dataset download is required.

import numpy as np
from rabitqlib import IvfIndex

rng = np.random.default_rng(42)
data = rng.standard_normal((500, 64)).astype(np.float32)
queries = rng.standard_normal((5, 64)).astype(np.float32)

# Assign vectors to five clusters and calculate their centroids.
cluster_ids = (np.arange(len(data)) % 5).astype(np.uint32)
centroids = np.stack(
    [data[cluster_ids == cluster].mean(axis=0) for cluster in range(5)]
).astype(np.float32)

index = IvfIndex(
    dim=64,
    max_elements=len(data),
    num_clusters=5,
    nbits=4,
    metric="l2",
)
index.build(data, centroids, cluster_ids)

ids, distances = index.search(queries, k=10, nprobe=5)
print(ids.shape, distances.shape)  # (5, 10) (5, 10)
print(ids[0])

Python bindings are also available for HnswIndex and SymqgIndex. See the Python examples for index construction, querying, and index persistence.

Build the Python bindings from source

Source builds require a C++17 compiler, CMake 3.15 or newer, and OpenMP. On Ubuntu or Debian:

sudo apt-get update
sudo apt-get install -y build-essential cmake libomp-dev
git clone https://github.com/VectorDB-NTU/RaBitQ-Library.git
cd RaBitQ-Library
python -m pip install .

C++ quick start

Requirements

• CMake 3.15 or newer
• a C++17 compiler with OpenMP support
• an x86-64 CPU supported by the selected kernels: most paths accept either AVX2 with FMA or AVX-512F/BW/DQ with FMA

CPU dispatch details

Most SIMD entry points select AVX-512 kernels when AVX-512F, AVX-512BW, and AVX-512DQ are detected; otherwise they use AVX2 when AVX2 and FMA are available. AVX-512 VPOPCNTDQ enables additional popcount kernels. The HNSW AVX-512 core path also checks for AVX2 and FMA. AVX-512 translation units are compiled with FMA enabled.

Clone and build the library and example programs:

git clone https://github.com/VectorDB-NTU/RaBitQ-Library.git
cd RaBitQ-Library

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel

Release builds enable native CPU tuning by default. To build a binary that can be moved between AVX2- and AVX-512-capable machines, configure with -DRABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF; the ISA-specific kernels will still be selected at runtime.

The index example executables are written to bin/. Their source code shows the complete indexing and querying workflows:

• IVF + RaBitQ
• HNSW + RaBitQ
• SymphonyQG

A separate RaBitQ quantization example demonstrates the lower-level quantizer API; it is provided as source and is not currently a CMake target.

To build and run the C++ test suite:

cmake -S . -B build -DRABITQ_BUILD_TESTS=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel
ctest --test-dir build --output-on-failure

GoogleTest is downloaded during test configuration. For a full benchmark on the GIST dataset, see example.sh. More detailed API and algorithm guidance is available in the documentation.

Choose the right building block

IVF and SymphonyQG use FastScan for batched estimates, while HNSW uses single-code AVX2 or AVX-512 kernels.

In typical workloads, 4-bit, 5-bit, and 7-bit quantization can achieve roughly 90%, 95%, and 99% recall, respectively, without reranking. Actual results depend on the dataset, index configuration, and search parameters.

Citation

If RaBitQ helps your research or system, please cite:

Jianyang Gao, Yutong Gou, Yuexuan Xu, Yongyi Yang, Cheng Long, and Raymond Chi-Wing Wong. “Practical and Asymptotically Optimal Quantization of High-Dimensional Vectors in Euclidean Space for Approximate Nearest Neighbor Search.” Proceedings of the ACM on Management of Data 3, 3, Article 202 (June 2025), 26 pages. https://doi.org/10.1145/3725413.

Contributing

Contributions are welcome. See the contributing guide for the build, formatting, pre-commit, and static-analysis workflows.

Acknowledgements

RaBitQ Library is developed by Yutong Gou, Jianyang Gao, Yuexuan Xu, Jifan Shi, and Zhonghao Yang. We thank Alexandr Guzhva, Li Liu, Chao Gao, Silu Huang, Jiabao Jin, Xiaoyao Zhong, and Jinjing Zhou for their valuable feedback.

License

RaBitQ Library is available under the Apache License 2.0.