PDF OCR

Self-hosted, GPU-accelerated OCR web app that converts scanned or image-based PDFs into searchable, editable documents.

Upload a PDF and get back:

• Searchable PDF — original layout preserved with an invisible, selectable text layer
• Markdown — structured plain text with tables and images extracted
• Word (.docx) — editable document ready for further editing

Built with PaddleOCR and PaddleOCR-VL for high-accuracy Chinese and multilingual document recognition. Runs entirely on your own hardware — no data leaves your network.

Features

• GPU-accelerated — CUDA-powered OCR and vision-language model inference via PaddlePaddle
• Multiple output formats — searchable PDF, Markdown (with images), and Word from a single upload
• Batch processing — upload multiple PDFs at once; jobs run in the background
• Multilingual UI — Traditional Chinese (繁體中文), Simplified Chinese (简体中文), and English
• Smart Chinese text handling — automatic Simplified-to-Traditional conversion for Traditional Chinese users; Simplified Chinese users get native output
• Real-time progress — live progress bar and status updates per page
• Self-hosted & private — everything runs locally in Docker; your documents never leave your server
• Single-container deployment — one docker compose up and you're running

Quick Start

Prerequisites: NVIDIA GPU, Docker, NVIDIA Container Toolkit

git clone https://github.com/cyanyux/pdf-ocr.git
cd pdf-ocr
docker compose up -d --build

Open http://localhost:5000 in your browser and start uploading PDFs.

First launch takes a few minutes to download model weights (~2 GB). Subsequent starts are instant.

Configuration

All settings are optional environment variables (set in docker-compose.yml):

Variable	Default	Description
PDF_OCR_DEVICE	auto	Force cpu to disable GPU
PDF_OCR_CUDA_VISIBLE_DEVICES	—	Select specific GPU (e.g. 0)
PDF_OCR_MAX_UPLOAD_MB	500	Max upload size per file (MB)
PDF_OCR_MODEL_IDLE_TIMEOUT	1800	Seconds before unloading idle models from VRAM
PDF_OCR_CLEANUP_INTERVAL	3600	Output file cleanup interval (seconds)
PDF_OCR_MAX_FILE_AGE	3600	Delete output files older than this (seconds)
PDF_OCR_DISABLE_HPI	false	Disable HPI/ONNX GPU acceleration
SECRET_KEY	auto-generated	Flask secret key for CSRF

API

All endpoints accept multipart file uploads and return JSON with job IDs for async polling.

Method	Endpoint	Description
POST	/api/ocr	Convert PDF to searchable PDF
POST	/api/markdown	Convert PDF to Markdown (zip with images)
POST	/api/word	Convert PDF to Word (.docx)
POST	/api/export	Markdown + Word from single VL pass
GET	/api/job/<id>	Poll job status and progress
POST	/api/cancel/<id>	Cancel a running job
GET	/api/download/...	Download completed output
GET	/api/health	Health check (device, GPU status)

Tech Stack

• OCR Engine: PaddleOCR 3.4 + PaddleOCR-VL-1.5
• GPU Framework: PaddlePaddle 3.2 with CUDA 12.6
• Backend: Python / Flask / Gunicorn
• Frontend: Vanilla JS single-page app
• Container: NVIDIA CUDA 12.6.3 + cuDNN on Ubuntu 24.04
• Document conversion: Pandoc, PyMuPDF, python-docx

Troubleshooting

Check service health:

curl http://localhost:5000/api/health

Verify GPU access inside container:

docker exec pdf-ocr nvidia-smi

GPU not detected?

1. Ensure NVIDIA Container Toolkit is installed
2. Restart Docker after driver updates: sudo systemctl restart docker
3. Rebuild: docker compose down && docker compose up -d --build

Security

This service has no built-in authentication. Do not expose it directly to the internet. Use a reverse proxy (Nginx, Caddy, Traefik) with access control for production deployments.

License

MIT