A vector database stores multidimensional vector representations and retrieves data by similarity to a query. In retrieval-augmented generation (RAG), it keeps contextual data separate from the language model, making sources easier to select and update. section 11.3 section 11.4
In a RAG workflow, vector databases help to:
- Find data relevant to a user query through similarity algorithms. section 11.4
- Pull multiple relevant files into a single query and provide context that can improve response accuracy. section 11.4
- Expose a control and privacy tradeoff when hosted platforms obscure storage choices or relevancy algorithms. section 10.2.3
RAG query
A RAG framework can query a vector database with a user question, retrieve several relevant files, and add them to the prompt. Retrieval errors can still supply inaccurate context and produce incorrect responses. section 11.4
