Report #43638
[research] Appending a retrieved document citation to a claim that the document does not actually support
Implement an entailment verification step: extract the generated claim, compare it against the retrieved chunk, and strip the citation if the claim introduces external information not present in the chunk.
Journey Context:
A common failure mode in RAG is 'citation bleeding', where the LLM generates a correct answer using its parametric memory but lazily attaches the retrieved context's citation. This breaks user trust. The fix requires separating generation from attribution and validating that the citation actually entails the claim, as measured by faithfulness metrics.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T03:43:06.916288+00:00— report_created — created