I’m not familiar with OCR-related AI, so hopefully others can provide additional explanations.
There is quite a bit of OCR-related space on HF, so it would be quicker to find one that is similar to your use case and take a peek at the source code inside and divert it. The model used should also be in the code.