SciSage

This is the official repository of SciSage.

📄 Paper: https://arxiv.org/abs/2506.12689

📊 Benchmark: https://huggingface.co/datasets/BAAI/SurveyScope

Features

• Multi-source Paper Extraction: Robust crawling from arXiv with fallback mechanisms
• Intelligent Analysis: AI-powered paper understanding and outline generation
• Structured Content Generation: Section-wise detailed analysis with proper citations
• Multi-model Support: Compatible with GPT-4, local models, and cloud services

Quick Start

1. Installation

git clone https://github.com/FlagOpen/SciSage.git
cd SciSage
pip install -r requirements.txt

2. Configuration

Model Setup

Edit core/model_factory.py:

llm_map = {
    "gpt-4": AzureChatOpenAI(...),
    "gpt-4o-mini": AzureChatOpenAI(...),
}

Pipeline Settings

Edit core/configuration.py and set your Default model you want to use.

3. Run pipeline

1. cd core
2. SET YOUR SEARCH API KEY, IF YOUR WANT YOUR YOUR LOCAL MODEL, SET THE ENVIRONMENT:
3. bash run.sh

4. Run local demo

bash run_demo.sh

5. RUN Local Server

export GOOGLE_SERPER_KEY="xxx"
## for general search: get from https://serpapi.com/dashboard
export SERPAPI_API_KEY="xxx"

## for Local LLM inference url, change to your local LLM server address
export LOCAL_LLM_HOST="xxx"

python3 server.py

Afater setup server, you can test the server

python3 client.py

Project Structure

SciSage/
├── benchmark/              # Paper extraction tools
│   └── get_paper_info.py   # Multi-source paper crawler for benchmark build
├── core/                   # Analysis pipeline
│   ├── main_workflow_opt_for_paper.py  # Main orchestrator
│   ├── paper_outline_opt.py            # Outline generation
│   ├── paper_poolish_opt.py            # Content polishing
│   ├── model_factory.py                # Model management
│   └── configuration.py                # Settings
└── eval/                   # Evaluation tools

License

MIT License - see LICENSE file for details.

Citation

If you use SciSage or SurveyScope in your research, please cite:

@misc{shi2025scisagemultiagentframeworkhighquality,
      title={SciSage: A Multi-Agent Framework for High-Quality Scientific Survey Generation},
      author={Xiaofeng Shi and Qian Kou and Yuduo Li and Ning Tang and Jinxin Xie and Longbin Yu and Songjing Wang and Hua Zhou},
      year={2025},
      eprint={2506.12689},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2506.12689}
}