A team wants to distinguish retrieval failures from generation failures. Which review data is most useful?

1. The answer alone with all context removed
2. Retrieved chunks, their ranks, the question, and the generated answer
3. Only the HTTP response byte count
4. Only the embedding model's marketing description