[symmetrist]

Tesseract can run locally and you can also use other open source OCR models which use neural networks. I am pretty sure it is possible to fine-tune open source LLMs on images and get reasonable results.