What actually reduces hallucinations in enterprise document QA — beyond better prompts?

We've hardened prompts as far as they'll go. Still getting plausible-but-wrong answers when documents have conflicting information or the answer simply isn't there. Our users can't tolerate false confidence — they need the model to say "I don't know" and cite sources when it does answer. What techniques beyond prompt engineering are teams using: confidence scoring, answer-grounding checks, retrieval filtering?