Retrieval-augmented generation demos well and degrades quietly. Chunking drift, stale indexes and ambiguous queries erode answer quality long before anyone files a ticket.
We treat RAG like any production system: a golden question set, automated grading for faithfulness and retrieval recall, and regression gates in CI so every index or prompt change is scored before it ships.
The result is a system whose quality is a number on a dashboard rather than a feeling — and a team that can improve it deliberately.