Context and knowledge

Vector database

stable definition
Machine-readable Download Markdown

Definition

A vector database stores vectors with identifiers and metadata, builds indexes over them, and answers similarity queries. Applications commonly store an embedding beside a pointer to source content rather than treating the vector as the source itself.

Search may be exact or approximate. Exact nearest-neighbor search compares against every eligible vector. Approximate nearest-neighbor indexes trade some recall for lower latency or memory use. Index family, distance metric, quantization, probe settings, and filters all affect which neighbors are returned.

Jeff Johnson, Matthijs Douze, and Hervé Jégou's 2017 FAISS paper described high-speed brute-force, approximate, and compressed-domain similarity search on GPUs, including billion-vector experiments. It supports the indexing mechanics used by many vector systems, but it did not define the database product category.

A database adds operational behavior around the index: persistence, updates, deletion, replication, namespaces, access control, and metadata filtering. Some general databases provide vector extensions, while some specialized services make vector search their primary interface. The product label says little about retrieval quality unless the embedding model and index settings are also known.

Deletion deserves special attention. Removing a source record should remove or invalidate its chunks, vectors, caches, and derived indexes. Otherwise semantic search can keep returning material that the system claims no longer exists.

Distinguish it from nearby terms

An embedding model creates vectors. A vector library may provide an in-process index without database operations. A vector database stores and searches vectors. RAG is an application pipeline that may use the database, and durable memory is an application policy about what state persists.

Check your understanding

A user loses access to a project, but precomputed vectors for its documents remain in a shared namespace. Which filter, deletion, cache, and audit controls must prevent the old content from being retrieved?