⭐ DyG-RAG: Dynamic Graph Retrieval-Augmented Generation with Event-Centric Reasoning (PDF)

DyG-RAG is the first event-centric dynamic graph retrieval-augmented generation framework specifically designed to capture, organize, and reason over temporal knowledge embedded within unstructured text data. By explicitly modeling the temporal dynamics of events as graph-structured representations, DyG-RAG facilitates enhanced temporal reasoning, yielding substantial accuracy improvement in temporal QA task.

The overall framework of DyG-RAG:

---

✨ Key Contributions

1. First Dynamic Graph Structure for Temporal Textual Knowledge: We introduce the first dynamic graph retrieval-augmented generation system that structures and stores temporal textual knowledge from an event-centric perspective.
2. Event Granularity with Explicit Temporal Encoding: We propose a novel dynamic event unit (DEU) granularity that explicitly embeds temporal information into the knowledge organization stage, enabling fine-grained temporal encoding.
3. RAG-Reasoning Integration for Downstream Tasks: The construction of DyG naturally supports the integration of retrieval-augmented generation with temporal reasoning for downstream temporal QA tasks, enabling the Time-CoT prompting approach.
4. Empirical Validation: We have verified DyG-RAG's superior performance across three distinct types of temporal question-answering datasets, demonstrating its effectiveness in real-world scenarios.

🚀 Quick Start

1. Environment Setup

cd DyG-RAG
conda create -n dygrag python=3.10
conda activate dygrag
# Install dependencies
pip install -r requirements.txt

2. OpenAI Configuration

Set your OpenAI API key:

# Option 1: API-Environment variable
export OPENAI_API_KEY="your_api_key_here"

3. Download Models

# Download required models
cd models
python download.py

Downloaded Models:

The script downloads two essential models:

1. Cross-Encoder Model:

   • Model name: cross-encoder/ms-marco-TinyBERT-L-2-v2
   • Local path: ./models/cross-encoder_ms-marco-TinyBERT-L-2-v2/
   • Size: ~67MB

2. NER Model:

   • Model name: dslim/bert-base-NER
   • Local path: ./models/dslim_bert_base_ner/
   • Size: ~1.2GB

4. Local Model Configuration[optional]

If you want to use local models instead of OpenAI, you can set up environment variables for the local embedding and LLM configuration (using BGE and Qwen as examples):

# Local BGE embedding model path (example: BGE-M3)
export LOCAL_BGE_PATH="/path/to/your/bge-m3"

# VLLM API service URL (example: local VLLM server)
export VLLM_BASE_URL="http://127.0.0.1:8000/v1"

# Model name for LLM (example: Qwen model)
export QWEN_BEST="qwen-14b"

5. Run Examples

# Basic usage with OpenAI
cd examples
python openai_all.py

# Using local models
python local_BGE_local_LLM.py

📁 Code Structure

DyG-RAG/
├── requirements.txt  # Dependencies
├── examples/         # Usage examples
├── graphrag/         # src codes of DyG-RAG
├── demo/             # TimeQA dataset example
├── datasets/         # three types of temporal QA datasets
└── models/           # Downloaded models

📝 TODO

1. Support for More Model Choices
2. Support for Diverse Vector Databases
3. Support for Diverse Graph Databases

🙏 Acknowledgments

This project is built upon the excellent work of nano-graphrag by Gustavo Ye. We extend our sincere gratitude to the original author for providing a simple, easy-to-hack GraphRAG implementation that serves as the foundation for our DyG-RAG system.