Measuring hallucination rates in RAG pipelines without ground truth labels

We've deployed a RAG pipeline serving ~500 daily queries across internal docs. The retrieval step (BM25 + dense embeddings) performs well by nDCG, but we're struggling to measure how often the generation step fabricates claims not supported by retrieved context. Manual annotation at scale is not feasible — we'd need someone to verify every claim in every response against the source documents. What automated or semi-automated approaches are you using? We're considering: 1. NLI-based fact-checking (premise=retrieved context, hypothesis=generated claim) 2. Self-consistency checks (sample k responses, flag claims not present in majority) 3. Citation verification (force the model to cite source spans, then verify the spans contain the claimed info) Has anyone benchmarked these approaches? What's the false positive rate like in practice?