[Feature Request] Add alternative LLM based OCR

Description


Follow up of

While Tesseract is "ok", LLM OCR offers significantly better quality for handwritten notes or complicated documents (like e.g. demonstrated in the paperless-gpt project). This is even works with a small locally hosted minicpm-v-4.5 model.

This blog post describes a technique that allows to use LLMs and also fulfill the OCR overlay requirement:


The related source can be found here:


The already merged Azure AI OCR is a step in the right direction, but for sensible documents it would be good to have to choice to process everything locally.

Other


_No response_