Nexuldra
// ai
Vector database
A vector database is a store built for embeddings: lists of numbers that capture the meaning of texts. Given a new text's numbers, it finds the most similar stored ones in milliseconds, using distance metrics rather than keyword matching. It is the retrieval engine behind semantic search, recommendation systems and every RAG deployment that answers from documents.
A vector database matters because classical databases match exact values, not meaning. Questions are paraphrases — nobody phrases a query in the manual's own words — and similarity search closes that gap. It also carries the metadata filters RAG needs in practice: retrieve only chunks the asking user is allowed to see, from a collection updated this month. Search quality, access control and freshness all live in the same query, which is why RAG deployments put a vector store at their centre.
Concretely: a few tens of thousands of document chunks — the size of a typical company knowledge base — fit comfortably in a vector database costing tens of euros a month, often less. Query latency sits in the tens of milliseconds, and inserting or deleting a document updates the index immediately, which keeps answers in sync with the source. Choosing the product is rarely the hard part; keeping the ingestion pipeline wired to wherever documents actually live is.
Frequently asked questions
- Do we need a dedicated vector database, or can Postgres do it?
- For most business knowledge bases, Postgres with the pgvector extension is enough: millions of vectors, filtered search, and one less system to operate. A dedicated vector database earns its place at larger scale or heavier load — until then, operational simplicity usually wins.
- What is the difference between a vector store and an index?
- The store keeps the vectors themselves; the index — HNSW or similar — is the structure that makes nearest-neighbour search fast instead of scanning everything. Most products bundle both, but the distinction matters when tuning: index parameters trade recall against speed, and small recall losses show up as retrieval misses, not errors.