Repair RAG Retrieval, Grounding & Citation Failures

The likely issue is a contract mismatch around RAG retrieval, chunking, filters, embeddings/index versions, ranking, context assembly, grounding, citations, stale content, access control, and evaluation. I would first define the invariant: A generated claim may cite a source only when the exact context fragment supplied to the model can be traced to an authorized current source and supports that claim; retrieval misses and unsupported citations must remain observable. Then I would trace the smallest owning boundary, apply or propose a bounded repair/review step, and add deterministic normal, boundary, failure and replay tests. The workflow preserves source access controls, privacy, document ownership, citation/provenance requirements, and product answer policy. If repository/workspace files are available, the findings can be mapped to exact paths and focused local validation; with partial evidence, the result stays an explicit diagnosis/evidence map with assumptions and no invented execution claims.