Vector Database

A vector database stores embeddings and retrieves entries using a mathematical similarity measure.

While relational queries commonly search for exact values, ranges or relationships, vector search answers questions such as: which stored entries are most similar to this query vector?

How the search works technically

An exhaustive similarity search compares the query vector with every stored vector. This provides an exact comparison, but may be too expensive for large datasets or strict response-time requirements.

Many systems use approximate indexing methods for large datasets. They reduce search time, but may miss individual optimal matches.

The index, parameters and filters should therefore be tested with realistic queries against the required retrieval quality, response time and cost.

The central building block for RAG

Vector databases are commonly used in RAG systems. Document passages are stored as embeddings and retrieved by similarity when a request arrives.

RAG is not limited to vector search alone. Depending on the data and requirements, full-text search, metadata filters and combined ranking methods may also be used.

What to look for when choosing one

Vector search is available in specialised products and as an extension to relational databases. The appropriate choice depends on data volume, query load, filtering needs, operating model and the existing platform.

Embeddings may encode sensitive information. Storage location, access controls, tenant isolation, deletion and contractual processing terms should therefore form part of the selection decision.