Feature Request: Offline OCR Support via Local Models

[Nitrogly]

Feature Request: Offline OCR Support via Local Models

Current Situation

Currently, OCR functionality in llm_tool_ocr_mistral relies exclusively on the Mistral AI cloud API, which requires an active internet connection and a paid API key. This makes OCR unavailable in air-gapped environments or for users who prefer a fully self-hosted, cost-free, private setup.

Requested Feature

Add support for offline OCR using locally hosted vision-capable models via Ollama or similar self-hosted backends — without any dependency on external paid APIs.

Suggested Local Models (Ollama-compatible)

All of the following support vision/OCR tasks and run fully offline:

• llava:7b
• llava:13b
• minicpm-v:8b
• qwen2.5vl:7b
• gemma3:12b

Thank you!

[ayushin]

have you tried any of these models to confirm the suitability to extract document data?