haha, I know that feeling! I worked on a RAG system for a pharmaceutical client and the hardest part was exactly this: You know, when everything looks fine, without error, but results are silently wrong!!! I think LLMs answering with full confidence on bad data is the most dangerous failure mode.