FAISS

FAISS (Facebook AI Similarity Search) is Meta’s open-source library for fast similarity search over dense vectors. Note the word library: FAISS is not a database. There’s no server, no persistence layer, no metadata filtering, no replication — just highly optimized C++ (with Python bindings) that builds an index over a pile of vectors in memory and answers nearest-neighbor queries at remarkable speed, on CPU or GPU.

FAISS matters because it’s the engine underneath much of the vector search world — many vector databases either use it directly or reimplement its algorithms. Learning it teaches you the actual mechanics: flat indexes for exact search, IVF for partition-based approximate search, HNSW graphs, and product quantization for compressing vectors when memory is tight. When you need raw speed on a single machine — offline batch jobs, research experiments, or search embedded inside your own service — FAISS skips every network hop and gives you full control over the speed-versus-recall tradeoff.

In practice you’ll pip install faiss-cpu (or faiss-gpu), embed your documents into a NumPy array, then build an index: index = faiss.IndexFlatL2(dim) for exact search or IndexIVFFlat / IndexHNSWFlat for approximate. Call index.add(vectors), then distances, ids = index.search(query_vectors, k). You handle everything a database would: mapping returned IDs back to documents, saving with faiss.write_index, and reloading on startup. LangChain and LlamaIndex both wrap FAISS as a zero-infrastructure local vector store.