This project is a document parsing tool based on DeepSeek-OCR. The tool can efficiently process PDF documents and images, providing powerful Optical Character Recognition (OCR) capabilities, supporting multi-language text recognition, table parsing, chart analysis, and many other features.
- Multi-format Document Parsing: Supports uploading and parsing documents in various formats such as PDF and images
- Intelligent OCR Recognition: Based on the DeepSeek-OCR model, providing high-precision text recognition
- Layout Analysis: Intelligently recognizes document layout structure and accurately extracts content layout
- Multi-language Support: Supports text recognition in multiple languages including Chinese and English
- Table & Chart Parsing: Professional table recognition and chart data extraction functionality
- Professional Domain Drawing Recognition: Supports semantic recognition of various professional domain drawings
- Data Visualization: Supports reverse parsing of data analysis visualization charts
- Markdown Conversion: Converts PDF content to structured Markdown format
PDF Document Parsing - Supports complex content including images and tables
| Multi-language Text Parsing | Chart & Table Parsing |
|---|---|
 |
 |
| Professional Domain Drawing Recognition (CAD, Flowcharts, Decorative Drawings) |
Data Visualization Chart Reverse Parsing |
|---|---|
 |
 |
- Operating System: Requires running on Linux system
- GPU Requirements: GPU ≥ 7 GB VRAM (16–24 GB recommended for large images/multi-page PDFs)
- Compatibility Note: RTX 50 series GPUs are currently not compatible, please use other GPU models
- Python Version: 3.10–3.12 (3.10/3.11 recommended)
- CUDA Version: 11.8 or 12.1/12.2 (must match GPU driver)
- PyTorch: Requires installing pre-compiled version matching CUDA
Execute the following script for one-click startup
# Install model weights and environment dependencies
bash install.sh
# Start services
bash start.shFirst, you need to download the DeepSeek-OCR model weights, which can be obtained from Hugging Face or ModelScope. The following example uses ModelScope:
pip install modelscope
mkdir ./deepseek-ocr
modelscope download --model deepseek-ai/DeepSeek-OCR --local_dir ./deepseek-ocrDownload the official project package
git clone https://github.com/deepseek-ai/DeepSeek-OCR.gitCreate a virtual environment to install model runtime dependencies
conda create -n deepseek-ocr python=3.12.9 -y
conda activate deepseek-ocrInstall Jupyter and corresponding kernel
conda install jupyterlab
conda install ipykernel
python -m ipykernel install --user --name dsocr --display-name "Python (dsocr)"Install PyTorch related components
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu118Install DeepSeek-OCR officially recommended vLLM version (v0.8.5+cu118-cp38-abi3-manylinux1_x86_64.whl)
pip install ./packages/vllm-0.8.5+cu118-cp38-abi3-manylinux1_x86_64.whlInstall project basic dependencies
cd ./DeepSeek-OCR/
pip install -r requirements.txtIf dependency conflicts appear during installation as shown in the image, you can ignore them as they won't affect actual operation.
Install flash-attn acceleration library.
pip install flash-attn==2.7.3 --no-build-isolationCreate a .env file in the project root directory and enter the model runtime address, for example:
MODEL_PATH=/root/autodl-tmp/deepseek-ocr
Start the backend
uvicorn main:app --host 0.0.0.0 --port 8002 --reloadInstall frontend dependencies
npm installStart the frontend
npm run devAfter successful startup, access the frontend address in your browser to use the tool.
We welcome contributions to the project through GitHub PR submissions or issues. We very much welcome any form of contribution, including feature improvements, bug fixes, or documentation optimization.
Scan to add our assistant, reply "DeepSeekOCR" to join the technical communication group and exchange learning with other partners.
