^ This.

I've found success generally better with Textract and then passing the OCR'ed text to OpenAI and using Pydantic to get it structured.