RAG is a data problem, not a model problem
Permission-aware retrieval over your truth beats a bigger context window. Notes from three knowledge-system builds.
Teams reach for bigger models when retrieval is what's actually broken.
The pattern that works
Hybrid search (lexical + vector) over a permission-aware index that syncs with the source of truth. If the user can't open the document, the agent must not quote it.
Three builds, one lesson
In all three knowledge systems we shipped, quality moved when we fixed chunking, freshness, and ACL filtering - not when we swapped models.