Your offline eval shows recall@5 of 92% on a labeled set, so the gold passage is in the context for nearly every question. Yet human graders mark 35% of answers wrong, and the wrong answers often ignore or contradict the gold passage. Where should the team invest next?

1. A Generation: tighten the prompt and context assembly, e.g., answer only from context, reorder chunks, and abstain when unsupported.
2. B Retrieval: switch to a higher-dimensional embedding model, since 92% recall still leaves room and better vectors will raise answer accuracy too.
3. C Retrieval: raise k from 5 to 25, so the generator sees more corroborating passages and is less likely to contradict the single gold passage.
4. D Chunking: shrink chunks so the gold passage is less diluted, since the model's failure to use it suggests the embedding of that chunk is weak.