LightOnOCR Service

FastAPI + RQ + Redis service that runs OCR on PDFs and images using the LightOnOCR-2-1B vision-language model. Jobs are queued, processed by workers, and results are returned as structured JSON that can be converted to Markdown.

Features

• Upload a PDF or image and queue OCR in Redis
• Background workers run the model and return per-page text
• Simple status endpoint for job polling
• Utility CLI to convert OCR JSON output to Markdown

API

• POST /process_file (multipart file upload)
• GET /jobs/{job_id}
• GET /health

Example:

curl -X POST "http://localhost:8000/process_file" -F "file=@sample.pdf"

Run Locally

1. Install dependencies:

pip install -r requirements.txt

1. Set environment variables (see .env.example).
2. Start Redis (local or Docker).
3. Start the API:

uvicorn app.main:app --host 0.0.0.0 --port 8000

1. Start a worker:

cd app
rq worker OCR_TASK_QUEUE

Docker

docker-compose up --build

JSON to Markdown Utility

Convert OCR JSON output to Markdown:

python -m app.utils.ocr_to_md path/to/ocr.json

Write to a specific file:

python -m app.utils.ocr_to_md path/to/ocr.json -o output.md

Environment Variables

• REDIS_HOST
• REDIS_PORT
• REDIS_QUEUE_NAME
• HF_TOKEN (optional, for Hugging Face access)

Model Attribution

This project uses the LightOnOCR-2-1B model from Hugging Face, which is described as a flagship OCR model and is licensed under Apache-2.0. The model card and details are available on Hugging Face.

Hugging Face model page