Source Citation

Explanation

Source Citation checks whether a generative AI system that answers questions or summarises documents — for example a retrieval-augmented (RAG) chatbot explaining a customer's loan eligibility or a tool summarising a policy document — actually backs up what it says with the source material it retrieved. This matters because in banking an ungrounded or hallucinated answer can mislead a customer, misstate a fee or product term, and breach transparency and explainability obligations; a reader must be able to trace every material claim back to a real source. It is measured by the Source Citation Rate, the share of material claims in a generated response that cite a source chunk (the specific retrieved passage the claim came from). To implement it, instrument generation so each retrieved chunk carries a stable identifier, require the model to attach a citation to each factual statement, and run a post-generation checker that splits the output into material claims and verifies each links to a supporting chunk; log the response, claims, cited chunk ids, and resulting rate as evidence. The threshold requires that >=95% of material claims cite a source chunk; falling below it triggers an alert and review — for instance withholding or flagging the answer, routing it to a human, or returning it for regrounding — consistent with RBI transparency and explainability expectations, NIST MEASURE 2.9 (model explained) and ISO Clause 7.4 documented-information requirements.