Installation

Platform requirements and source installation for RaBitQ Library 0.7.2. For the first Python example, see Quick Start.

Requirements

Platform	CPU baseline	Python wheel targets
Linux x86-64	AVX2 and FMA	CPython 3.11–3.14
Linux ARM64 (AArch64)	NEON	CPython 3.11–3.14
Windows x86-64	AVX2 and FMA	CPython 3.11–3.14
macOS 14+ ARM64 (Apple Silicon)	NEON	CPython 3.11–3.14

Source builds require a C++17 compiler and CMake 3.20 or newer. Full ANN/Python builds also require OpenMP; core/FastScan can disable it with RABITQ_USE_OPENMP=OFF. Windows uses MSVC (Visual Studio 2026 with the Desktop development with C++ workload and CMake 4.2+ for its generator). Apple Silicon uses AppleClang and an external OpenMP runtime such as Homebrew libomp.

Linux ARM64 source builds and repaired wheels run on native AArch64 CI. Linux ARM64 wheels carry manylinux_2_27_aarch64 and manylinux_2_28_aarch64 tags.

CPU dispatch details

On x86-64, optional AVX-512 kernels require AVX2, FMA, and AVX-512F/BW/DQ; MSVC builds also require AVX-512VL/CD. Popcount-specific kernels additionally require AVX-512 VPOPCNTDQ. Detection checks CPU features and OS support for the required register state. Missing AVX-512 features select the AVX2 backend. On ARM64, the build excludes x86 kernels and uses NEON and portable scalar implementations. Standard FastScan requires AVX2/FMA, a supported AVX-512 backend, or NEON; scalar fallbacks do not remove that requirement. A pure-core build can use scalar kernels without the FastScan CPU baseline.

Python wheels

python -m pip install "rabitqlib>=0.7.2"

For changes from 0.6.0, see the 0.7.0 upgrade notes.

Wheels target the platforms above and require no compiler or CMake. Every wheel bundles its OpenMP runtime (GNU OpenMP on Linux, LLVM OpenMP on macOS, and the MSVC vcomp140 runtime on Windows), so wheel users do not need a separate OpenMP installation. Release wheels disable native CPU tuning and select supported kernels at runtime.

Python editor support

Version 0.7.2 includes type declarations (__init__.pyi) and a py.typed marker for editor completion and static type checking. Select the Python environment where rabitqlib is installed. Index and clustering docstrings describe parameters, defaults and returned arrays; use help(IvfIndex) or help(RaBitQKMeans) after importing those classes. Array shapes and parameter ranges are still checked at runtime.

Get the source

For Python source builds or C++ use, obtain the 0.7.2 checkout once. With Git installed, run:

git clone --branch v0.7.2 --depth 1 https://github.com/VectorDB-NTU/RaBitQ-Library.git
cd RaBitQ-Library

Run the source-build commands below from this repository root. If you already have the matching checkout, use it directly.

Build the Python bindings from source

Source builds require Python 3.11 or newer, a C++17 compiler, CMake 3.20 or newer, and OpenMP. First get the source, then on Ubuntu or Debian run:

sudo apt-get update
sudo apt-get install -y build-essential cmake libomp-dev
python -m pip install .

On Apple Silicon, use a native ARM64 Python environment. From the repository root:

brew install cmake ninja libomp
python -m pip install . -Ccmake.define.OpenMP_ROOT="$(brew --prefix libomp)"

On Windows, install the build tools listed above, then run python -m pip install . from the repository root. For a portable source build on any supported platform, also pass -Ccmake.define.RABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF.