Making RAG admit when it's guessing: source-grounded hallucination checks

Short summary

SWIRL 5 mitigates RAG hallucination by verifying each generated claim against its cited sources using entailment checks, flagging unsupported claims inline before users see them. The approach segments answers into atomic claims, extracts cited text spans, and detects gaps like citations that are topically relevant but don't contain the quoted data. Rather than claiming to eliminate hallucination, source-grounded verification tells users when answers are uncertain and points to problematic sentences.

• •Entailment checks verify if citations actually support generated claims
• •Atomic claim segmentation and citation span extraction catch hallucination gaps
• •Inline flagging informs users when the system is uncertain rather than hiding it

Generated with AI, which can make mistakes.