Agentic DeepSeek-OCR wrapper for turning PDFs and images into Markdown with a customizable vLLM pipeline and Gradio UI.

DeepReader

DeepReader is an agentic reading toolkit that couples DeepSeek-OCR with opinionated defaults for running single-document or batch OCR. It streamlines image/PDF ingestion, produces Markdown accompanied by figure crops and layout previews, and exposes knobs for both CLI and Gradio workflows.

Project Layout

• images/: Sample page images for quick smoke-tests.
• docs/: Input PDFs for full-length papers.
• outputs/: Generated Markdown, annotated images, and layout PDFs.
• DeepSeek-OCR-master/DeepSeek-OCR-vllm/: vLLM-powered runtime (default entry points).

Environment Setup

• download the vllm-0.8.5 whl

conda create -n deepreader python=3.12.9 -y
conda activate deepreader
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 \
  --index-url https://download.pytorch.org/whl/cu118
pip install vllm-0.8.5+cu118-cp38-abi3-manylinux1_x86_64.whl
pip install -r requirements.txt

Optional extras:

• pip install flash-attn==2.7.3 --no-build-isolation (faster attention if supported).

Configuration Strategy

DeepSeek-OCR-vllm/config.py reads all defaults from environment variables, making it easy to swap inputs, outputs, prompts, or GPU settings without editing code.

export DEEPREADER_INPUT_PATH="$PWD/docs/paper.pdf"
export DEEPREADER_OUTPUT_PATH="$PWD/outputs/paper_run"
export DEEPREADER_PROMPT='<image>
<|grounding|>Convert the document to markdown.'
export DEEPREADER_PROMPT_TEMPLATE=document
export DEEPREADER_MODE=gundam
export DEEPREADER_CUDA_VISIBLE_DEVICES=0
export DEEPREADER_GPU_MEM_UTIL=0.8
export DEEPREADER_KEEP_MODELS_LOADED=1

GPU tip: the default vLLM config assumes ≈10 GB of free VRAM. Tune DEEPREADER_GPU_MEM_UTIL down if you’re memory-constrained.

Gradio Interface

Launch an inte