[Template only, I am not the container author/maintainer]

 

Project: https://github.com/icereed/paperless-gpt

Support: https://github.com/icereed/paperless-gpt/issues

Registry: https://hub.docker.com/r/icereed/paperless-gpt

 

paperless-gpt seamlessly pairs with paperless-ngx to generate AI-powered document titles and tags, saving you hours of manual sorting. While other tools may offer AI chat features, paperless-gpt stands out by supercharging OCR with LLMs-ensuring high accuracy, even with tricky scans. If you're craving next-level text extraction and effortless document organization, this is your solution.
Description of container variables: https://github.com/icereed/paperless-gpt?tab=readme-ov-file#configuration

There is an update for paperless-gpt, now adding Mistral API as option. I updated the container, but there are no options to fill the new variables in the template. Is it possible to update this?

Edited July 7, 2025Jul 7 by MrNossi

Posting the steps I took to successfully use the "GLM-OCR" model for OCR duties in paperless-gpt with an ollama backend.

1) Make sure you get glm-ocr from ollama.com. I'm using the "bf16" version, not the "q8_0", since VRAM is not an issue for me. I would avoid using any other .gguf, as an example from huggingface.com, with the ollama backend as I think those other versions were assembled to be used on other backends and could be slightly different.

2) I'm using context lengths between 8k and 64k with this model. Anything smaller or larger than that results in poor output or the model crashing for me.

8k = 8192

16k = 16384

32k = 32768

64k = 65536

3) I changed the ocr prompt in paperless-gpt to the following:

-----------------------------------------------------------------------------------------------

Extract text from the provided document/image while preserving all original formatting and layout. Use precision mode on. Output a text file with these requirements:

• Maintain original spacing between paragraphs and lines.

• Preserve special characters and mathematical notation (use ... for equations).

• Do not add any metadata, comments, or modifications beyond what appears in the document.

• For unclear text, leave as-is without correction.

• Ignore non-text elements unless they contain readable text.

Output should be a clean text file containing only the document's content.

-----------------------------------------------------------------------------------------------

The model has 3 modes, text recognition, table recognition, and figure recognition. I believe the output that we would be striving for is text since that is what the "Content" field in paperless-ngx supports. As an example, when you get html output, I believe the model is trying to recreate a table via table recognition mode.

The above ocr prompt, for me, does a good job of steering the model to focus on text extraction mode.

I've successfully used this on images and pdfs.

you are right! Thank you. Was experimenting with long financial tables and went for table-mode, that was the reason for the <td> spam likely. And randomly increasing the context window solved that "Asian signs" problem.

However, since for my use case I need the model to also describe graphical elements from and pictures time to time I went back to qwen3-vl. That one still produces random md garbage on rotated tables with almost no content (zeros) and is stupid slow (like 90min per 30 pages on hard documents), but still the best i could achieve by now. Would be awesome if you could configure paperless-gpt to fire multiple requests against the LLM at once. I am running vllm-rocm as back end.

According to the paperless GPT GitHub Page there should bei Support for docling as the OCR Provider. There are noch Options for the docling Server url in the template and I was Not successful adding it myself. Any suggestions or experience on how getting it to work?