Building a Hallucination-Detection Pipeline for Production

Leadership wants a way to catch hallucinations in your RAG assistant's production answers before they reach a customer complaint, without hand-reviewing every conversation (you get ~50,000 conversations/day). Design a layered hallucination-detection pipeline: what runs on every answer, what runs on a sample, and how the two feed into each other and into the offline gold set over time. Be explicit about what each layer catches that the others miss.