Definition
A vector database stores vectors together with identifiers and metadata, then provides indexes and query operations for similarity search. Depending on the product, it may also support filters, multiple vector fields, payload updates, replication, and operational controls.
Its core role is storing and searching vectors. Embedding generation and application-level retrieval policy are separate concerns, even when a product integrates those capabilities.
Simple example
A documentation pipeline stores one vector for each passage along with its source URL, tenant ID, language, and document version. At query time, the application sends a query vector plus a tenant filter. The database returns nearby passages that satisfy that filter.
Why it matters
Scanning every stored vector becomes expensive as a collection grows. A vector database packages indexing, persistence, filtering, and query execution behind an operational interface. That can make large similarity workloads practical.
One important nuance
Vector search does not require a dedicated database system. PostgreSQL with pgvector can store and search vectors beside relational data, while Qdrant is a dedicated vector database. The choice depends on scale, filtering, consistency, and the infrastructure you already operate. Matching dimensions do not mean vectors share a compatible embedding space. If incompatible embeddings must stay separate, record and enforce the embedding contract in the schema, ingestion pipeline, or application.