RAG that survives legal review: citations, permissions, and evaluation
AI transformation dies in the CISO’s inbox when retrieval-augmented generation cannot prove where an answer came from. Production RAG is an access-control problem first.
Dr. Priya Nandakumar · Head of AI Systems · 9 min · August 28, 2026
Demo RAG is a vector store and a prompt. Production RAG is a permissions graph, a citation contract, and an evaluation harness that fails closed. If your assistant can see a document the employee cannot, you do not have a knowledge bot. You have a data leak with a friendly UI.
Chunking strategy determines whether citations are usable. Too large and the model quotes the wrong paragraph. Too small and you lose the policy context legal requires. Hybrid search plus a reranker beats “embed everything and hope,” especially on policies, contracts, and engineering runbooks.
Evaluation is not a vibe check. Build a golden set from real tickets and questions, score faithfulness and refusal quality, and re-run it on every index refresh. Drift is normal. Undetected drift is how a confident wrong answer reaches a customer.
Privacy sandboxes (private endpoints, PII redaction, retention windows) are what make AI transformation fundable. The model is interchangeable. The governance around retrieval is the moat.