Vector Database

A vector database treats embeddings as first-class, indexable data. Each record usually contains a vector plus metadata such as a document ID, timestamp, tenant, or access-control label. A query is embedded into the same space, then compared with stored vectors using cosine distance, Euclidean distance, or an inner product.

The common operations are insertion, nearest-neighbor search, metadata filtering, and updates to the index. In retrieval-augmented generation, document chunks are embedded once and searched at request time so the language model receives relevant context instead of relying only on its parameters.

Large collections rarely use an exact comparison against every vector. Approximate nearest-neighbor structures such as HNSW and IVF trade a small amount of recall for much lower latency and memory cost. The central caveat is that vector distance is only a proxy for relevance. A nearby vector can be fluent, topical, and still fail to answer the question.