Skip to content

Vector store

A vector store keeps embeddings and returns nearby items according to a similarity function. Useful records also keep the original text, source location, permissions, tenant, and other metadata.

A query embedding is compared with stored passage embeddings. The store returns five passage IDs, their similarity scores, and the metadata needed to fetch and cite the source text.

A vector store does not make retrieved content correct, current, or authorized. It also does not replace lexical search for exact identifiers.

Apply tenant and document-permission filters inside retrieval. Filtering results after search can leak metadata and reduce the useful result count.