No direct recommendation for that use case, but one strategy I've heard being used and that works with complex documents (or where hallucinations are Very Bad™ - like invoice processing) is using multiple techniques and models at once in a quorum approach. For example, direct ingestion of PDFs into Gemini, OCR and ingestion of text, plus perhaps using another model like GPT. If they all agree on a fact, you're (probably) good. If not, it can be bumped up to human correction.