RAG vs fine-tuning for policy-heavy domains

Policy documents change. Fine-tuned weights do not, unless you operate a retraining loop you can defend. For SNAP/Medicaid-style guidance, retrieval plus citations was the path that kept lawyers and program owners in the conversation.

That is not a universal rule. Classification, routing, and short canned intents may belong in a smaller model. Long-form “what does this regulation mean?” usually belongs in retrieval.

Where have you seen RAG fail in production — stale indexes, bad chunking, or over-confident answers when retrieval is empty? Add a comment with a concrete failure mode if you have one.