A trustworthy AI citation connects one material claim to an exact passage in a versioned source and records whether that passage really supports the claim. A URL list is not enough.
A RAG result is relevant only when the caller may see it. Authorization must constrain candidate retrieval before text reaches reranking, prompts, caches, logs, or model providers.
A RAG index is a derived projection with document versions, chunker and embedding identities, tombstones, ACL changes, atomic generations, and a measurable freshness SLO.
Reranking more candidates helps only while relevant evidence enters the pool and changes final order. Measure quality, latency, cost, duplicates, and permission-safe failure by query slice.
Your tenant has enough documents, but an HNSW query returns too few. Compare real plans, iterative scans, and an exact fallback without weakening access checks.