We have an eval (and I want to add more), but basically depending on the model image to text quality varies a lot if it is not grounded.
The good news though is that with PDF we can do text extraction in a lossless way and then only lean on the LLM to improve it if you want to gold plate. Should have something next week.