What Is a Vector Database?

A vector database is a storage and search system built for embeddings: it holds millions to billions of content vectors and, given a query vector, returns the most similar ones in milliseconds. It is the retrieval backbone of RAG pipelines — when an AI engine "finds" relevant content, a vector search across an index like this is usually what happened.

The mechanics that matter

Purpose-built systems (Pinecone, Weaviate, Qdrant, Milvus) and extensions like pgvector for PostgreSQL all implement the same core operation: approximate nearest neighbor (ANN) search, typically via graph indexes such as HNSW. "Approximate" is deliberate — exact search over billions of vectors is too slow, so indexes trade a sliver of recall for speed. Each stored vector carries metadata (source URL, date, section), and queries can filter on it: only chunks from the last year, only from documentation.

What retrieval mechanics imply for content design

Reading the pipeline backward yields concrete content rules:

• Top-k is a hard cliff. The database returns a fixed number of chunks — often 10 to 50 — and everything else is invisible to the answer, however good. Being almost the closest match pays nothing; distinctive, query-matching passages are what clear the cut.
• Metadata filters are real gates. Recency filters make honest dateModified values and genuinely updated content mechanically valuable, not just reassuring.
• The unit of competition is the chunk. Your page enters the index as independent passages; each must survive on its own meaning, the argument developed under semantic chunking.
• Redundancy across phrasings helps. Covering a topic in multiple naturally-varied passages gives you multiple points in vector space, matching multiple query formulations.

Scope note

Public answer engines like Perplexity or ChatGPT Search run hybrid retrieval stacks — web-scale keyword indexes plus semantic layers — rather than a single off-the-shelf vector store. But vector search principles govern the semantic half everywhere, and thousands of enterprise assistants your buyers use internally are literal RAG-over-vector-DB systems reading your public content. Designing for them is part of the same GEO practice.