Keep semantics connected to records
An embedding helps compare semantic similarity, but rarely replaces the structured record behind it. A relational database lets you join vector results to permissions, document owners, timestamps and application state.
Filter before exposing results
Retrieval-augmented generation still needs authorization. The application should restrict which documents may be considered and verify that returned data belongs to the correct tenant or user, even when a semantic index spans many records.
Plan for freshness and evaluation
Track source changes, embedding model versions, deletion semantics and relevance testing. A working similarity query is not enough to establish reliable product behavior.
Where UnDataly fits
pgvector is on UnDataly's private-alpha roadmap. Treat code samples and UI concepts as previews until the access, isolation and performance gates have passed.
Frequently asked questions
Are these UnDataly features available now?
Guides explain general architecture patterns. The linked UnDataly managed features are planned or in development unless otherwise marked.