Tell me about a time you took end-to-end ownership of a retrieval-augmented generation (RAG) pipeline problem when the path forward was unclear and action was needed quickly. What signals told you there was a real issue, how did you decide what to change in retrieval, ranking, or grounding, and what tradeoffs did you make under time pressure? Walk me through the outcome and what you would do differently now.