Abstention

Explanation

Abstention checks that the system declines to answer — saying "I don't know" or escalating — when retrieval returns nothing relevant or only low-confidence matches, instead of guessing; this is the primary guardrail against hallucination in RAG, summarization and chat use cases, and matters acutely in BFSI where a confidently invented answer about, say, KYC requirements or loan eligibility can cause real customer harm and regulatory exposure. It is measured by the Abstention Rate on Low-Confidence/Empty Retrieval (the metric and calcMethod are the same): of all cases where retrieval is empty or below the confidence cutoff, the fraction in which the system correctly abstained. Implement it by defining an explicit confidence/relevance threshold on retrieval scores, gating generation behind that check, and emitting a deterministic refusal or human handoff when the gate fails; log every query, its retrieval scores, the abstain/answer decision, and the trigger reason so behaviour is auditable. The threshold is strict — 100% abstention on empty or low-confidence retrieval — and this is an Immediate-priority control, so any answered response in a low-confidence/empty case is a breach that must be alerted on and corrected, since even a single unfounded answer defeats the control's purpose.