cuVS is a GPU-accelerated library for approximate nearest neighbors (ANN) search and clustering, derived from the RAPIDS RAFT library. It aims to simplify GPU usage for vector similarity search and clustering tasks, targeting developers and researchers working with large embedding datasets for applications like semantic search, generative AI, and recommender systems.

How It Works

cuVS leverages high-performance GPU primitives from the RAPIDS RAFT library to implement state-of-the-art ANN and clustering algorithms. It focuses on providing efficient implementations of methods like CAGRA for ANN search and cuSLINK for single-linkage agglomerative clustering, enabling faster index builds and low-latency, high-throughput search operations.

Quick Start & Requirements

• Install: conda install -c conda-forge -c nvidia -c rapidsai cuvs or pip install cuvs-cu12 --extra-index-url=https://pypi.nvidia.com for CUDA 12.
• Prerequisites: NVIDIA GPU with CUDA support.
• Resources: Requires GPU memory proportional to the dataset size for indexing and search.
• Docs: Library documentation, Build and Install Guide, Getting Started Guide, Code Examples.

Highlighted Details

• Provides GPU-accelerated implementations of CAGRA (ANN search) and cuSLINK (clustering).
• Offers Python, C++, C, and Rust APIs for integration.
• Designed for low-latency, high-throughput search and fast index builds.
• Algorithms will be removed from RAFT in v24.12, with cuVS becoming the primary source.

Maintenance & Community

• Part of the broader RAPIDS ecosystem.
• Community support available via RAPIDS Community.
• Source code and issue tracker available on GitHub.

Licensing & Compatibility

• License: Apache 2.0.
• Compatible with commercial use and closed-source applications.

Limitations & Caveats

• Primarily targets NVIDIA GPUs; CPU interoperability is mentioned but not detailed as a primary feature.
• As a newer library consolidating RAFT features, long-term maintenance and feature parity with RAFT's older implementations should be monitored.