A vector database is a library of embeddings: it stores millions of vectors and returns the “nearest neighbors” in milliseconds. ANN techniques like HNSW keep the operation scalable.
In RAG architectures this is the external memory: your knowledge base is embedded, stored here and fetched as relevant chunks at answer time. The well-known names: Pinecone, Weaviate, Qdrant, Milvus and pgvector.
Selection note: if your stack is already PostgreSQL, pgvector is often enough; a dedicated vector DB arrives when scale or hybrid-filter needs turn serious.
