ANN recall hides RAG relevance gap

This is the right complaint: vector search can look healthy while RAG still underperforms, because retrieval quality and evidence usefulness are not the same metric.

• –Recall@k, nDCG, and MRR help, but they still stop short of telling you whether the chunk was actually the best evidence for the answer.
• –Once ANN recall is decent, hybrid search plus reranking usually buys more than another round of HNSW/IVF tuning.
• –Measure the whole chain with retrieval-process relevance, groundedness, response completeness, and answer-level relevance.
• –For debugging, log the exact prompt context and run ablations or chunk swaps to see whether the generator really depends on the top-ranked chunk.