At instances where data accuracy is of paramount importance, i think a hybrid route of non-llm ocr for data parsing and LLMs for structured data extraction is the safe passage to tread on. Seen better results for LLMWhisperer(OCR)[1] and Latest Gemini.