A vector database stores embeddings along with the original text or an ID and any metadata, and is optimized for one question: which stored vectors are closest to this one? It usually uses approximate nearest-neighbor search, trading a little exactness for much faster results across large collections.
It is the storage layer behind semantic search and most Retrieval-Augmented generation setups. Some are dedicated products; some are extensions to a regular database, so a separate system is not always needed, especially at small scale.
Useful features are metadata filtering (for example only search one product's docs), keeping vectors in step when source content changes, and being able to delete or update entries. Stale vectors are a quiet cause of wrong answers, so plan for refreshing them.
Related: Knowledge Base, Document Chunking, AI Agent Memory.