---
id: libcuvs-cu12
version: "26.8.1"
license: Apache-2.0
license_treatment: permissive
maintenance: active
---

# libcuvs-cu12 — cuVS: Vector Search on the GPU (C++)

License: permissive · Maintenance: active · Downloads: 114.5K/mo

## What it is and what it does

libcuvs-cu12 is a GPU-accelerated library for vector search and clustering built on NVIDIA CUDA. It implements state-of-the-art approximate nearest neighbor algorithms and clustering primitives designed to run on NVIDIA GPUs, with the goal of accelerating similarity search on dense vector embeddings. The library is part of the RAPIDS ecosystem and provides APIs for multiple languages.

The package is used for semantic search tasks—including retrieval-augmented generation, recommendation systems, and image/text search—as well as data mining operations like clustering, visualization, and k-NN graph construction. It requires CUDA 12 and depends on libraft-cu12, librmm-cu12, nvidia-nccl-cu12, and nvidia-nvjitlink-cu12, meaning it is only usable on systems with compatible NVIDIA GPUs and the CUDA toolkit installed.

Use it for:

- Build fast approximate nearest neighbor indexes for semantic search in retrieval-augmented generation and generative AI applications
- Accelerate recommendation systems by computing similarity between user embeddings and item vectors on GPU
- Perform large-scale clustering and visualization tasks on GPU-resident data
- Construct k-NN graphs from dense vectors for graph analysis and downstream machine learning
- Enable low-latency, high-throughput vector similarity queries in production retrieval systems

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

GPU-accelerated vector search and clustering library providing approximate nearest neighbor algorithms for similarity search on NVIDIA hardware.

Yes, if you have NVIDIA GPU hardware and need fast vector search or clustering. The library is actively maintained, has no known vulnerabilities, and the Apache-2.0 license is permissive. Install friction is moderate due to CUDA and GPU library dependencies, but these are expected for GPU-accelerated workloads. Not suitable for CPU-only environments.

## Install

pip install libcuvs-cu12
uv add libcuvs-cu12
poetry add libcuvs-cu12

## Installing libcuvs-cu12

Before you install: Medium install friction due to CUDA toolkit and multiple NVIDIA GPU libraries (libraft-cu12, librmm-cu12, nvidia-nccl-cu12, nvidia-nvjitlink-cu12) as runtime dependencies. Active maintenance with recent release and steady repository activity.

License in practice: Apache-2.0 permissive license allows commercial and private use with minimal restrictions, making it suitable for integration into proprietary applications.

Quickstart:

pip install libcuvs-cu12

from libcuvs-cu12 import neighbors

# Build approximate nearest neighbor index on GPU-resident data
index_params = neighbors.IndexParams()
index = neighbors.build(index_params, dataset)

Requires NVIDIA GPU with CUDA 12 support and CUDA toolkit installed; Python >=3.11

Verify before relying:

- Whether pre-built wheels cover all target architectures beyond aarch64 and x86_64
- Specific CUDA version compatibility requirements beyond the cu12 variant designation
- Performance characteristics or throughput benchmarks for different dataset sizes
- Whether the Python API is available in this package or requires separate installation

## Package facts

- License: Apache-2.0 (permissive)
- Python support: supports_current
- Install friction: medium
- Maintenance: active
- Downloads: 114.5K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

gpu vector search, approximate nearest neighbors gpu, cuda clustering algorithms, vector similarity search gpu, gpu-accelerated embeddings, nearest neighbor index cuda, semantic search gpu, gpu-accelerated, vector-search, rapids

[View on SkillFed](https://skillfed.io/packages/libcuvs-cu12) · [View on PyPI](https://pypi.org/project/libcuvs-cu12/)