Measuring hallucination rates in RAG pipelines without ground-truth labels

We've got a production RAG system over internal documentation (~50k pages). We need to quantify hallucination rates for a quarterly review, but manually labeling outputs is not scalable. Approaches we've explored: - Self-consistency: run the same query 5x with different temperature seeds, measure disagreement - NLI-based entailment: check if the generated answer is entailed by the retrieved chunks - Fact extraction: pull atomic claims from the answer, verify each against the source corpus The NLI approach gives the cleanest signal but struggles with nuanced technical content. Self-consistency is cheap but conflates genuine ambiguity with hallucination. Has anyone found a middle ground that doesn't require labeled datasets? Particularly interested in approaches that work well over technical/corporate docs rather than Wikipedia-style text.