Feature Request: Add OCR Backend Support for Local Document Processing

[20yuto20]

Summary

Propose adding an OCR (Optical Character Recognition) backend to enable local document text extraction capabilities within Docker Model Runner.

Motivation

• Expand Docker Model Runner beyond text generation to include vision/document processing
• Enable privacy-focused local OCR without cloud dependencies
• Leverage existing model distribution and scheduling infrastructure

Proposed Implementation

1. Create new OCR backend following existing patterns in pkg/inference/backends/
2. Integrate with popular document AI, e.g., layoutLMv3, Donut, and et cetera
3. Support common image formats (PNG, JPEG, PDF)
4. Expose OCR functionality through OpenAI-compatible API endpoints

Technical Considerations

• Follow existing backend interface in pkg/inference/backends/llamacpp/llamacpp.go
• Leverage model distribution system for OCR model downloads
• Integrate with resource management for memory allocation
• Support both CPU and GPU acceleration where available

Questions for Maintainers

• Preferred document AI models?
• API endpoint design preferences?
• Model packaging/distribution strategy?

Comment

I would be very grateful in my work if I could easily test document AI; OCR locally!