LLama 3.2-vision without second stage OCR

[giuliofrey]

Hi,

Thank you for providing this service. I see that LLaMA 3.2-vision is already quite capable and might not need a second stage OCR. To my knowledge, the Python wrapper currently does not allow for this.

Am I doing something wrong?

[choinek]

Hi!

The service is not performing the second stage of OCR.

We have two phases (or stages as you referred to):

1. Parameter: "strategy" (initial phase) – This phase converts the document into a unified format that can be processed by AI

   • Currently, we only have OCR strategies for PDFs and images. However, we are working on Docling support, which will enable us to handle almost all types of text documents (e.g., word files) via our API – see [feat] Add docling support #54

2. Parameter: "model" (final phase) – This phase uses AI to structure the document into its final format

In the near future, we will create documentation to provide an architectural visualization of the entire process. :)

But please clarify what do you mean by "Python wrapper currently does not allow for this."
I'm not sure what you're referring to. Are you encountering any errors? If so, could you share more details?

[pkarw]

I think you mean you can't setup the dynamic prompt for Ollama based models, if you'd be able to you could potentially do the extraction and remodeling within single step right?

[giuliofrey]

Yes exactly!