What is Vector database?

There is no single formal specification for vector databases, and the cited page is one vendor's description. It is the Qdrant overview, and the page itself never gives a formal definition of the term. It describes the idea through vector search: embedding models convert text, images or audio into numeric vectors, similar items land near each other in a high-dimensional space, and a query is converted into a vector so the engine can return the K closest matches.

Qdrant organizes data into collections, which are named groups of points and the set that a search runs within. A point is the basic record and pairs a vector with optional metadata called the payload. Points are identified by 64-bit integers or UUIDs. For each named vector, the collection defines the dimensions and a distance metric. A point can hold several named vectors, dense or sparse.

Search uses an HNSW index, a graph structure that lets the engine move through similar vectors quickly. Payload indexes extend that graph so filter conditions apply during the vector search in one pass. Qdrant also supports keyword-style lexical search and combines it with semantic search in hybrid retrieval.

Other products can differ in details, so this page should not be read as describing every vector database.