Mitigating hallucinations in RAG

Hallucinations in RAG have three root causes: (1) generator hallucination despite correct context: reduce via constrained decoding, low temperature, system prompts that demand citations; (2) generator hallucination due to bad retrieval: the context is wrong; (3) generator hallucination due to weak context: context is right but ambiguous, model invents to fill gaps. Mitigation: (a) cite-then-answer prompts forcing the model to ground specific claims to specific spans; (b) claim-level faithfulness scoring (split answer into atomic claims, verify each against context); (c) self-consistency / multi-sample voting; (d) re-rank with cross-encoders; (e) upstream curation: bad retrieval is more often the cause than generator bugs, so measure retrieval Recall@k separately. Senior nuance: hallucination evaluation is harder than generation quality because the truth boundary depends on doc updates; LLM-as-judge must be calibrated with a held-out human-labeled set.