Retrieval Provenance

Explanation

Retrieval provenance checks, for RAG systems, whether every answer the system produces records the specific source chunks that were retrieved and fed to the model to generate it. It matters because a RAG answer with no logged provenance cannot be verified or audited - if a customer-facing assistant cites a policy or a regulator questions an output, the bank must be able to show exactly which document passages the answer was grounded in, and detect when an answer was produced from no or irrelevant context. The metric and calculation method are the Retrieval Chunk Provenance Log itself: a persisted record, per answer, of the retrieved chunks (their identifiers, source documents, and the passages used). To implement it, instrument the retrieval step so that for each generated answer it writes a structured provenance entry linking the answer to the chunk ids, document references, and retrieval scores that contributed, stored in tamper-evident, queryable logs that can be replayed during audit or dispute. The threshold is that 100% of answers log chunk provenance, meaning no answer may be returned without its provenance recorded; any answer missing a provenance log is a breach that should trigger investigation and a fix to the logging path so that provenance capture is mandatory before an answer is served. Aligns with RBI Transparency and Explainability, NIST MEASURE 2.8 traceability, and ISO Clause 7.5.3 control of documented information.