A vector database is a database management system built specifically to store, index, and query high-dimensional vectors at scale. Unlike a relational database that retrieves records by exact matches on keys, a vector database retrieves records by similarity, finding the stored vectors closest to a query vector in meaning.
Its core jobs are to ingest vectors along with their metadata, build and maintain an index that organises them geometrically, and answer nearest-neighbour queries with low latency across millions or billions of vectors. To do this efficiently it relies on approximate-nearest-neighbour algorithms, support for metadata filtering, and infrastructure for scaling, all tuned for the particular demands of vector search.
Vector databases are the infrastructure layer behind modern AI applications. They serve as the retrieval engine for retrieval-augmented generation, the long-term memory for AI agents, the backbone of semantic search, and the core of recommendation systems. What distinguishes a purpose-built vector database from bolting vector search onto a traditional database is its optimisation for these workloads — fast filtered similarity search, real-time updates, and scale — which becomes decisive as vector search moves from a feature to a foundation.