Accuracy

Explanation

This accuracy control checks whether a RAG retriever surfaces all of the genuinely relevant source passages for a query, not just some of them — the coverage side of retrieval quality. It matters in banking because a partial retrieval can omit the one clause, fee schedule, or regulatory caveat that changes the correct answer, producing a response that looks complete but is materially wrong. It is measured by Retrieval Recall@K: of all the passages that are truly relevant to a query, the fraction that appear within the top K returned by the retriever. No calculation method beyond this metric is provided, so report standard recall-at-K against a labelled relevance set rather than inventing a formula. Implement it with a maintained gold set of queries and their full set of relevant passages, periodically run the live retriever, and log recall, the K used, and the relevant passages that were missed as evidence; instrument it at the retrieval stage before generation. Operate against the target of greater-than-or-equal-to 0.90; investigate below 0.85 by tuning chunk size, embeddings, or K, and remediate below 0.80, where too much relevant context is being missed for answers to be reliable.