I Added a Verify Layer to My Local RAG to Catch Hallucinations. It Caught Me Being Wrong Twice About My Own Corpus

Short summary

The author built a verify layer for local RAG that decomposes answers into atomic claims and checks each against source passages using LLM-as-judge. Testing on 8 controlled claim pairs achieved 100% hallucination detection with zero false positives, assuming perfect retrieval. The implementation revealed that right numbers in wrong contexts fool both models and human evaluators alike.

• •Verification layer decomposes RAG answers into claims, checking each against source passages via LLM-as-judge
• •8-pair controlled benchmark showed 100% false claim detection with zero false positives under perfect retrieval
• •Key insight: right numbers in wrong contexts bypass verification—a hallucination type neither model nor human initially caught

Generated with AI, which can make mistakes.