How do I tell the underlying model to not produce numbers and fractions in the latex format?

[UntouchedWagons]

I've got paperless-gpt set up like so:

services:
  paperless-gpt:
    image: icereed/paperless-gpt:v0.25.1
    container_name: paperless-gpt
    restart: unless-stopped
    environment:
      PAPERLESS_BASE_URL: "http://webserver:8000"
      PAPERLESS_API_TOKEN: Loremipsumdolorsitamet
      LLM_PROVIDER: "ollama"
      LLM_MODEL: "gemma4:e4b"
      OLLAMA_HOST: "http://ollama:11434"
      OLLAMA_CONTEXT_LENGTH: "8192" # Sets Ollama NumCtx (context window)
      TOKEN_LIMIT: 1000 # Recommended for smaller models
      LLM_LANGUAGE: "English" # Optional, default: English
      OCR_PROVIDER: "llm" # Default OCR provider
      VISION_LLM_PROVIDER: "ollama" # openai, ollama, mistral, or anthropic
      VISION_LLM_MODEL: "gemma4:e4b" # minicpm-v (ollama) or gpt-4o (openai) or claude-sonnet-4-5 (anthropic/claude)
      AUTO_OCR_TAG: "paperless-gpt-ocr-auto"
      AUTO_TAG: "paperless-gpt-auto"
      MANUAL_TAG: "paperless-gpt-manual"
      PDF_OCR_TAGGING: "true"
      PDF_OCR_COMPLETE_TAG: "paperless-gpt-ocr-complete"
      PDF_UPLOAD: "false"

And this seems to be working well but I'm having an issue where the underlying model produces numbers and fractions in the latex format. Instead of 28 ⅜ or 28 3/8, $28 \frac{3}{8}''$ is produced. Can anything be done about this?

[darekwojciechowski]

Yes, this is fixable from the OCR prompt — but your compose file is missing the volume that makes prompt edits stick.

Add this to the paperless-gpt service:

    volumes:
      - ./prompts:/app/prompts

Create the folder next to your compose file and restart. On first run the defaults get copied in.

Then open the paperless-gpt web UI, go to Settings, and edit the OCR prompt. Add a line like:

Transcribe exactly what is printed, as plain text. Never use LaTeX, math mode, or markup of any kind. Write fractions the way they appear on the page, for example 28 3/8, not \frac{3}{8}.

Templates reload as soon as you save, so no restart needed to test.

If it still slips through, that's the model rather than the prompt. gemma4:e4b is small, and small vision models drift into LaTeX because so much of their training data with fractions was LaTeX. Try a dedicated vision model for VISION_LLM_MODEL and keep gemma for the text tasks — the OCR prompt is the one that matters here, since the fractions are coming out of the vision pass, not the tagging pass.

To check which stage is doing it, run one document and look at the raw OCR text in paperless before any tagging happens. If the \frac is already there, it's the vision prompt you need to edit, not the general one.