AI Automated Paper-to-Code Workflow

This project implements an automated pipeline to convert research papers (specifically in the domain of Planning and Combinatorial Optimization) into structured data and code mappings. It uses a multi-agent LLM approach to digitize, analyze, and ground the papers into their corresponding code repositories.

🚀 Features

• Multi-Modal Ingestion: Converts PDF page images into structured Markdown using qwen-vl-plus.
  • New: Supports multi-threaded OCR (parallelism = 3 by default) for faster processing.
• Deep Reasoning: Uses kimi-k2-thinking to extract mathematical formulations, constraints, and objective functions into standard LP (Linear Programming) formats.
• Code Grounding: Maps mathematical symbols and logic to actual source code implementations using qwen-max.
• Structured Storage: Embeds knowledge using text-embedding-v4 and stores it in a local ChromaDB vector database.

🏗 Architecture

The workflow consists of three main processors managed by a facade:

1. IngestionProcessor: Handles OCR and layout analysis.
2. AnalysisProcessor: Extracts semantic logic and mathematical models.
3. GroundingProcessor: Binds the abstract logic to concrete code implementation.

🛠 Setup

1. Prerequisites:

   • Python 3.10+
   • DashScope API Key (Alibaba Cloud) for Qwen models.
   • Moonshot API Key for Kimi models.

2. External Data Source: Clone the awesome-ml4co repository parallel to this project folder:

   cd ..
   git clone https://github.com/Thinklab-SJTU/awesome-ml4co.git
   cd ai_paper_workflow

3. Installation:

   conda env create -f environment.yml

4. Environment Variables: Set your API keys in the environment:

   export DASHSCOPE_API_KEY="your_dashscope_key"

🏃 Usage

1. Data Preparation

To download sample papers from the ML4CO dataset:

python scripts/prepare_ml4co_data.py

This will download papers and repos into the scripts/data directory.

2. Run Single Pipeline

To run the workflow on a specific paper and repo:

python -m src.main --repo_path "path/to/repo" --paper_images "page1.jpg" "page2.jpg" ...

3. Batch Processing

To run the workflow on randomly selected downloaded papers:

python run_on_downloaded.py

📂 Output

Intermediate results are saved in output/{paper_id}/:

• full_paper.md: The digitized markdown text.
• analysis.json: Extracted LP formulations.
• grounding.json: Symbol-to-Code equivalence mappings.

🤖 Models Used

• Vision: qwen-vl-max / qwen-vl-plus
• Reasoning: kimi-k2-thinking
• Code Understanding: qwen-max
• Embedding: text-embedding-v4