Generative AI Hallucinations: Is there ANY robust way to prevent them reliably in factual content?

I'm developing an AI agent for legal document review and the occasional factual 'hallucination' is a deal-breaker. We've tried RAG, stricter prompting, and even smaller, domain-specific models, but occasional fabrication still slips through. Is there a truly robust, perhaps novel, approach to practically eliminate these for high-stakes applications? Really need some actionable insights here.