A vector database stores documents as numerical representations of their meaning (embeddings), so a search finds what is semantically close, not just what shares keywords. Ask "how do I get reimbursed?" and it retrieves the refund policy even if the document never uses the word "reimbursed". It is the memory layer behind most RAG systems.
In practice it is one component, not a product decision to agonise over: the retrieval quality depends far more on how documents are cut and indexed — work we cover in the RAG chatbot guide.