What happens when the LLM is confidently wrong? How do you debug a RAG chatbot giving confident but wrong answers?