RAG Hallucination Reduction: Faithfulness and Grounding Metrics

Hallucination types: intrinsic (contradicts source), extrinsic (unverifiable from source). RAGAS metrics: faithfulness (claims in answer grounded in context), answer relevancy (answer addresses question), context recall (relevant docs retrieved), context precision (retrieved docs are relevant). FaithfulAI suite: FactScore, SAFE, SelfCheckGPT. Mitigation: retrieval-augmented generation, chain-of-thought with citations, RLHF from factuality feedback. RAG faithfulness: each claim should be...

• rag
• hallucination
• faithfulness