Evaluating Retrieval and Answers Separately
One accuracy number cannot say whether a RAG failure came from retrieval, context selection, or generation. Score each stage, audit the judges, and diagnose by stage.
One accuracy number cannot say whether a RAG failure came from retrieval, context selection, or generation. Score each stage, audit the judges, and diagnose by stage.
Retrieval can succeed and the answer can still be a confident guess. Select evidence, check citations, resolve conflicts, abstain when the evidence is missing, and verify.
Sparse and dense retrievers fail on different queries. Learn how to fuse them, where a reranker helps or hurts, how to rewrite queries without losing exact terms, and how to benchmark each stage.
A retrieval index is derived data. Give every chunk an exact source span, name each index build, update approximate indexes without silent recall loss, and test for stale evidence.
A model’s weights are a lossy, frozen snapshot of the world. Retrieval is how you stop relying on them.