Diagnosing a High-Recall, Low-Faithfulness Eval Result

Your offline RAG eval on a 200-query labeled gold set reports recall@5 = 0.95 (the correct chunk is present in the top-5 for 95% of queries) but faithfulness = 0.60 (an LLM-judge check finds only 60% of generated claims are actually supported by the retrieved chunks the answer cites). Given these two numbers together, what does this result most directly indicate, and where should the fix be targeted first?

A. Retrieval is doing its job — the right chunk is usually present — but generation isn't reliably grounding its claims in that chunk; the fix belongs in generation and grounding checks, not in re-indexing. B. The embedding model is stale and should be swapped for a newer one to fix the low faithfulness score. C. Chunk size is too large, diluting the vector index's precision and causing the faithfulness gap. D. The BM25 index has fallen out of sync with the dense index and needs to be rebuilt.