Built for similarity search
A vector database stores data converted into numeric sequences by embedding and finds items closest in meaning at speed. It came to prominence alongside generative AI, and particularly with the spread of RAG.
How it differs from an ordinary database
A conventional database excels at exact matches and conditional queries. A vector database excels at finding what is most similar.
Locating the chunk closest in meaning to a question among millions of documents would be far too slow by brute force, so these systems use indexing techniques designed to find near neighbors efficiently.
Where it is used
- RAG: the retrieval layer for internal document and support assistants
- Image search: finding photographs or products with a similar feel
- Recommendation: matching users and items by proximity of taste
- Long-term memory for agents: recalling past exchanges by meaning
Worth knowing
Alongside dedicated products such as Pinecone, Qdrant, and Chroma, adding vector search to an existing database such as PostgreSQL is also common. Individuals rarely touch one directly, but knowing that embeddings plus a vector database sit behind giving an AI organizational knowledge makes deployment articles and RAG explanations far easier to read.
Vector search is strong on similarity and weak on exact conditions such as dates and amounts, so production systems commonly pair it with conventional search in a hybrid arrangement. It comes up in every discussion of improving RAG accuracy.
Embeddings, vector databases, and RAG as a set covers the foundation of most AI system discussion.