A vector database stores embeddings and serves nearest-neighbour search at scale. The retrieval step in RAG is almost always a vector database query under the hood.
Leading options in 2026: Pinecone (managed), Weaviate (open-source plus managed), Qdrant (open-source), Chroma (lightweight, local-first), pgvector (PostgreSQL extension, popular for small projects). Choice depends on scale, operational preference, and whether you want to colocate vectors with relational data.