Integrates with the MTEB leaderboard for embedding model evaluation and selection in the GraphRAG implementation.
Provides core agent orchestration and tool chaining capabilities for implementing ReAct and TS-ReAct agent patterns.
Enables graph-based agent workflow construction and state management for implementing ReAct and TS-ReAct agent patterns.
Provides LLM capabilities through GPT-4o, GPT-4, and GPT-3.5 Turbo models as backbone options for running ReAct-based agents.
Powers graph neural network operations through PyTorch Geometric for GraphRAG-based agent implementations.
⚠️ Important Notice ⚠️
As the paper is under review, all contents in this repository are currently not permitted for reuse by anyone until this announcement is removed. Thank you for your understanding! 🙏
1. Overview & Objectives
This repository contains the complete implementation, experimental data, and supplementary results for the paper ××× developed by XXX University in China, and .
Pending publication, the code is shared under a restrictive license. Once the paper is accepted, the repository will transition to a MIT license. Please contact the corresponding author for any inquiries regarding academic use during the review period.
2. Videos of agents operation
2.1 Operation of the developed prototype
↓↓↓ A snippet of using the developed prototype to run the TS-ReAct-based agents driven by GPT-4o
↓↓↓ A snippet of updating the tool kit in the prototype
The full video to showcase the prototype and tool kit updating can be found in:
2.2 Operation of agents based on ReAct pattern
↓↓↓ A snippet of running the ReAct-based agents driven by GPT-4o, GPT-4, and GPT-3.5 Turbo.
The full video can be found here ()
↓↓↓ A snippet of running the ReAct-based agents driven by Qwen2.5, Deepseek-V3, Gemma-2, Llama-3.1, and Mixtral MoE.
The full video can be found here ()
2.3 Operation of agents based on TS-ReAct pattern
↓↓↓ A snippet of running the TS agent based on TS-ReAct pattern.
The full video can be found here ()
↓↓↓ A snippet of running the ReAct agent based on TS-ReAct pattern.
The full video can be found here ()
3. Repository Structure
4. Acknowledgments
This work heavily relies on excellent open-source projects, including but not limited to:
LangGraph & LangChain
Hugging Face MTEB leaderboard
NetworkX, PyTorch Geometric, and numerous LLM providers (OpenAI, Anthropic, Qwen, Llama, etc.)
We are deeply grateful to all contributors of these foundational work.
5. How to Reuse This Repository
5.1 Importing the Lifeline Recovery Tool Set
Copy all tool definition files from
tools/into your target agent directory.Import the tools using the standardized registry pattern shown in the example notebooks.
5.2 Running Baseline ReAct Agents
Directory:
agents_reAct/Supports 8 different LLMs (GPT-4o, Claude-3, Llama-3.1-405B, Qwen2.5, etc.)
Ready-to-run scripts with configuration YAMLs
5.3 Running the Proposed GraphRAG + MCP Agents
Directory:
agents_graphRAG_MCP/Same 8 backbone LLMs
Includes GraphRAG index construction scripts and MCP search configurations
5.4 Running the Interactive Prototype
Directory:
prototype/Dynamic tool registration/hot-reloading
Web-based GUI + terminal interface
Supports on-the-fly addition of new recovery actions