RGL is a friendly and efficient Retrieval-Augmented Generation on Graphs (RAG on Graph) library for AI researchers, providing seamless integration with DGL and PyG while offering high-performance graph retrieval algorithms, many of which are optimized in C++ for efficiency.
✅ Seamless Integration – Works smoothly with DGL and PyG
⚡ Optimized Performance – C++-backed retrieval algorithms for speed
🧠 AI-Focused – Tailored for GraphRAG research and applications
🔗 Scalability – Handles large-scale graphs with ease
- Homepage: https://github.com/PyRGL/rgl
- Documentation: https://rgl.readthedocs.io
- Paper Access: RGL: A Graph-Centric, Modular Framework for Efficient Retrieval-Augmented Generation on Graphs
- DGL: https://www.dgl.ai/pages/start.html
- PyG (Optional): https://pytorch-geometric.readthedocs.io/en/latest/install/installation.html
pip install rglcd clibs
./build_linux.shfrom sklearn.feature_extraction.text import CountVectorizer
from rgl.utils import llm_utils
from rgl.node_retrieval.vector_search import VectorSearchEngine
from rgl.graph_retrieval.retrieve import retrieve
# Define nodes.
node_data = [
{"title": "Deep Learning for Graph Data", "classes": "AI, ML"},
{"title": "Introduction to Neural Networks", "classes": "ML"},
{"title": "Graph Neural Networks in Practice", "classes": "ML, Graph"},
{"title": "Advanced Topics in Machine Learning", "classes": "ML"},
{"title": "Reinforcement Learning in Robotics", "classes": "Robotics, ML"},
{"title": "Large Language Models and Reasoning", "classes": "AI, NLP"},
{"title": "Bayesian Methods in Deep Learning", "classes": "ML, Bayesian"},
{"title": "Computer Vision for Autonomous Vehicles", "classes": "CV, AI, Robotics"},
]
# Define edges. The two lists represent connections: the first list is the source nodes and the second list is the destination nodes.
edges = [
[1, 1, 1, 1, 0, 0, 2, 2, 3, 3, 4, 5, 6],
[0, 2, 3, 6, 2, 3, 3, 4, 6, 7, 7, 7, 0],
]
# Convert to an undirected graph by adding reversed edges.
src, dst = edges
src, dst = src + dst, dst + src
# Prepare title features and initialize vector search engine.
titles = [paper["title"] for paper in node_data]
vectorizer = CountVectorizer()
paper_feats = vectorizer.fit_transform(titles).toarray()
vector_search_engine = VectorSearchEngine(paper_feats)
# Input query paper title and retrieve similar papers (anchors).
query_paper_title = "Vision Transformers for Traffic Sign Recognition"
query_vector = vectorizer.transform([query_paper_title]).toarray()
retrieved_indices = vector_search_engine.search(query_vector, k=3)[0][0]
anchors = [node_data[idx] for idx in retrieved_indices]
# Retrieve subgraph from anchors.
anchor_indices = [node_data.index(paper) for paper in anchors]
subgraph_nodes = retrieve(src, dst, anchor_indices)
# Construct prompt with the query title and relevant paper information.
relevant_paper_str = "\n".join(
[f"Title: {node_data[node]['title']}, Classes: {node_data[node]['classes']}" for node in subgraph_nodes]
)
prompt = (
"Given the paper title: '{}'\n\n"
"And relevant paper information:\n{}\n\n"
"List the classes that are most relevant to the query paper."
).format(query_paper_title, relevant_paper_str)
print("\n=== Prompt Sent to Model ===\n{}".format(prompt))
# Query the LLM to obtain paper classification output.
output = llm_utils.chat_openai(prompt, model="gpt-4o-mini")
print("\n=== RAG Paper Classification Output ===\n {}".format(output))Output:
=== Prompt Sent to Model ===
Given the paper title: 'Vision Transformers for Traffic Sign Recognition'
And relevant paper information:
Title: Graph Neural Networks in Practice, Classes: ML, Graph
Title: Deep Learning for Graph Data, Classes: AI, ML
Title: Advanced Topics in Machine Learning, Classes: ML
Title: Computer Vision for Autonomous Vehicles, Classes: CV, AI, Robotics
List the classes that are most relevant to the query paper.
=== RAG Paper Classification Output ===
The classes that are most relevant to the query paper 'Vision Transformers for Traffic Sign Recognition' are:
1. Computer Vision (CV)
2. Artificial Intelligence (AI)
3. Machine Learning (ML)
These classes are particularly relevant because traffic sign recognition is a task within the domain of computer vision and typically involves machine learning techniques, including advanced models like Vision Transformers.
We recommend you to get started with some demo.
After running the demo, please visit http://127.0.0.1:7860 in your browser to access the interface.
@misc{li2025rglgraphcentricmodularframework,
title={RGL: A Graph-Centric, Modular Framework for Efficient Retrieval-Augmented Generation on Graphs},
author={Yuan Li and Jun Hu and Jiaxin Jiang and Zemin Liu and Bryan Hooi and Bingsheng He},
year={2025},
eprint={2503.19314},
archivePrefix={arXiv},
primaryClass={cs.IR},
url={https://arxiv.org/abs/2503.19314},
}