FAISS
FAISS (Facebook AI Similarity Search) is a library for efficient similarity search and clustering of dense vectors. It is developed mainly at Meta’s Fundamental AI Research group. It is a library, not a database: it builds and searches indexes in your own process, and storage, metadata filtering and serving are your job.
Repository snapshot (GitHub API, 2026-11-19)
| Item | Detail |
|---|---|
| Repo | facebookresearch/faiss, C++ with Python wrappers, MIT licence |
| Created | 2017-02-07 |
| Activity | last push 2026-11-19; latest release v1.16.0 (2026-11-10) |
| Size | 41,819 stars, 4,656 forks |
What it does (README)
- Searches sets of vectors of any size, including sets that do not fit in RAM. Vectors are identified by an integer and compared by L2 distance or dot product; cosine similarity works as a dot product on normalised vectors.
- Compact methods (binary vectors, quantisation codes) keep only compressed vectors, trading precision for scale: billions of vectors in memory on one server. Graph methods such as HNSW and NSG add an index structure on top of raw vectors for faster search. See ann-index-algorithms.
- GPU implementations of the most useful algorithms, single or multi-GPU, usable as drop-in replacements for CPU indexes; optional NVIDIA cuVS and AMD ROCm support.
- Installs from conda packages (
faiss-cpu,faiss-gpu,faiss-gpu-cuvs) or builds with cmake; the only hard dependency is a BLAS library. - Includes evaluation and parameter-tuning code.
Where it fits
FAISS is the index engine under many other tools: you use it directly for in-process semantic search over vector-embeddings, or let a framework wrap it for RAG. Choose a vector database or a search engine such as opensearch or azure-ai-search when you also need persistence, filters, access control, replication and an API; see vector-databases-compared.
Sources
- FAISS README and GitHub API (2026-11-19)