NVIDIA cuVS

NVIDIA cuVS is an open-source library for vector search and clustering on the GPU, published by NVIDIA under the Apache 2.0 license. It provides approximate and exact nearest neighbor algorithms, vector compression and clustering, usable directly or through databases and libraries that integrate it. The latest GitHub release listed is v26.10.00.

What It Is

cuVS is a developer library that simplifies using GPUs for vector similarity search, preprocessing and clustering. It offers both core building blocks for composing new algorithms and end-to-end algorithms. It is built on the RAPIDS RAFT library of machine learning primitives.

Workflow and Language Support

Developers build an index (for example a CAGRA graph index) over a dataset of vectors and then search it for nearest neighbors. The README shows this workflow in Python, C++, C and Rust, and the documentation also has API references for Java and Go. Installation is available through conda, pip or tarball, with a Docker path for building the standalone C library.

Algorithms and Use Cases

The documentation covers indexing methods such as brute-force, CAGRA, IVF-Flat, IVF-PQ, NN-Descent, ScaNN and Vamana, clustering such as K-Means, Gaussian mixture models, single-linkage and spectral clustering, and preprocessing such as quantizers and PCA. Documented use cases include semantic search for RAG and agentic AI, recommender systems, image, text, audio and molecular search, and data mining tasks like k-NN graph construction. The docs also describe multi-GPU support, dynamic batching and a benchmarking tool, cuVS Bench. The documentation says the benefits include fast index build, parameter tuning and building on GPU then deploying on CPU.