A lightweight tool to convert PDF documents into structured Markdown and JSON. It supports table repair, image captioning, heading recognition, and structured data extraction.
Docling2md is a document understanding and conversion tool built on the Docling framework and integrated with vision-language models (VLMs) and large language models (LLMs). Key features include:
- Converting PDFs into structured Markdown
- Automatically extracting headings and paragraph structure
- Using a VLM to generate image captions
- Repairing tables when table image recognition fails (table slicing + VLM reconstruction)
- Emitting structured JSON for downstream tasks such as knowledge graphs and search indexes
git clone https://github.com/your-username/Docling2md.git
cd Docling2mdpip install -r requirements.txtExtract the poppler archive to the project root or install it system-wide.
OPENAI:
api_key: "<your-deepseek-api-key>"
base_url: "https://api.deepseek.com"
model: "deepseek-chat" # default chat model for DeepSeek
max_concurrency: 10 # maximum concurrent requests for the chat model
VLM:
api_key: "<your-dashscope-api-key>"
base_url: "https://dashscope.aliyuncs.com/compatible-mode/v1"
model: "qwen2.5-vl-72b-instruct" # default VLM model
max_concurrency: 3 # maximum concurrent VLM requests
OCR:
enabled: true
POPPLER:
path: "D:/your_path_to/poppler/Library/bin" # path to extracted poppler binariesinput_pdf_path = Path("D:/Docling2md/your_path_to_pdf_folder")python pdf2md.pyOr embed convert_pdf_to_markdown_with_images() into your own pipeline.
- Markdown file:
output/<pdf_hash>/<hash>.md - JSON file:
output/<pdf_hash>/<hash>.json - Page images:
output/<pdf_hash>/page/page-*.png - Table and image files:
output/<pdf_hash>/*.png
graph TD
A[PDF Document] --> B[DocumentConverter]
B --> C[Structured Document]
C --> D[Element Iteration]
D --> E[Text Processing]
D --> F[Table Processing]
D --> G[Image Processing]
E --> H[LLM Classification]
F --> I[Table Structure Analysis]
G --> J[VLM Captioning]
H --> K[Markdown Output]
I --> K
J --> K
K --> L[Markdown File]
K --> M[JSON Metadata]
K --> N[Extracted Images]
- PDF layout parsing: Uses Docling to extract tables, images, and text elements
- Table image repair: Supports automatic slicing + VLM reconstruction for problematic tables
- Image captioning: VLM generates short captions and embeds them in Markdown
- Text classification: Uses DeepSeek to decide whether text is a heading or paragraph
- Structured output: Produces unified JSON that includes fields like
type,level,page_number, andbbox
.
├── pdf2md.py
├── config.yaml
├── prompt/
│ ├── VLM_prompt.py # used for image caption prompts
│ ├── text_type_prompt.py # used to decide heading vs paragraph
│ ├── text_repair_prompt.py # used to repair text without spaces
│ └── table_repair_prompt.py # used to repair problematic table images
├── output/
│ └── <pdf_hash>/
│ ├── page/
│ ├── *.png
│ ├── *.md
│ └── *.json
└── ...
| tD(on) | - | 20 | - | ns | V_DS=400V, V_GS=0 - 12V, I_D=3A, R_G=30Ω |
| ------- | - | -- | - | -- | ---------------------------------------- |
| tR | - | 7 | - | ns | V_DS=400V, V_GS=0 - 12V, I_D=3A, R_G=30Ω |
| tD(off) | - | 80 | - | ns | V_DS=400V, V_GS=0 - 12V, I_D=3A, R_G=30Ω |
| tF | - | 6 | - | ns | V_DS=400V, V_GS=0 - 12V, I_D=3A, R_G=30Ω |
If you have suggestions or issues, please open an issue or submit a PR.
This project is licensed under the MIT License.
This project integrates Qwen-VL and DeepSeek models and is intended for academic research and technical validation only.