NVIDIA/cuvs

What it solves

cuVS simplifies the use of GPUs for vector similarity search and clustering. It addresses the need for high-performance information retrieval of semantically similar items from large datasets of multimedia embeddings, which is critical for modern AI applications like RAG and recommender systems.

How it works

Built on top of the RAPIDS RAFT library, cuVS provides state-of-the-art GPU-accelerated implementations of approximate nearest neighbors (ANN) and clustering algorithms. It allows users to build indices and perform searches across multiple languages (Python, C++, C, and Rust) and supports interoperability between GPU-built indices and CPU deployment.

Who it’s for

Developers and data scientists working with dense vectors and embeddings who need low-latency, high-throughput search and fast index building for semantic search, data mining, or graph analysis.

Highlights

• Multi-language support: APIs available for Python, C++, C, and Rust.
• High performance: Optimized for NVIDIA GPU architectures to ensure low latency and high throughput.
• Versatile use-cases: Supports semantic search (RAG, image/text/audio search), data mining (clustering, visualization), and k-NN graph construction.
• Interoperability: Ability to build indices on GPU and deploy them on CPU.