Accepted by ICDE2026 !!!

ClueRAG: Towards Accurate and Cost-Efficient Graph-based RAG via Multi-Partite Graph and Query-Driven Iterative Retrieval.

Note: This is a new version, the oceanbase augmented version, and its README.md is coming soon.

Introduction

​Clue-RAG​​ is a novel ​​Retrieval-Augmented Generation (RAG)​​ framework that enhances LLM question answering by efficiently retrieving and integrating structured knowledge from a ​​multi-partite graph index​​.

🔍 ​​Key Features​​:

✅ ​​Multi-partite Graph Index​​ – Combines ​​text chunks, knowledge units, and entities​​ for fine-grained semantic retrieval

🚀 ​​Hybrid Extraction​​ – Reduces LLM token usage while maintaining high-quality knowledge disambiguation

🔍 ​​Q-Iter Retrieval​​ – Query-driven iterative search with ​​semantic matching & constrained graph traversal​​

🔥 Zero Token Consumption: Complete graph construction and retrieval without any LLM calls in Clue-RAG-0.0.

🏗️ Clue-RAG 's ​​offline indexing​​ (Clue-Index) transforms raw text into a structured knowledge graph through two stages: 1. ​​Hybrid Extraction​​ 2. Graph Construction​​. Q-Iter follows an intelligent 3-phase approach: 1. Entity Anchoring 2. Iterative Retrieval 3. Final Re-ranking.

Performance

Overall performance of RAG solutions. The best and second-best results among the ten competitor solutions (excluding the Clue-RAG series) in each column are highlighted in bold and underlined, respectively. Notebly, it outperforms the strong baseline KETRAG (KDD 2025) by 21.53%/29.16% in average F1/Acc., while also surpassing another competitive baseline SIRERAG (ICLR 2025) by 5.75% in average F1/Acc.

Setup Environment

1. Create a Python 3.10 Environment

conda create -n cluerag python=3.9 -y
pip install -r requirements.txt

2. Install spaCy

python -m spacy download en_core_web_trf

or download the en_core_web_trf-3.8.0 model from spaCy model releases, then

pip install en_core_web_trf-3.8.0.tar.gz

3. Quick Start

⚙️ Configure parameters: All hyperparameters are located in utils/config.py

📊 Data: The benchmark dataset used in this project is located in the data/ directory. They are multi-hop QA dataset widely adopted by the Graph-based RAG research community.

🚀 Run the main script: (from the project root directory):

python main.py

📂 Output Directory: When running experiments, results are automatically saved with this naming pattern: outputs/{dataset_name}/{select_metric}_{alpha}/

Example:
If you test with:

• dataset_name: 2wikimultihopqa
• select_metric: COSINE
• alpha value: 1

Your results will be saved in:
outputs/2wikimultihopqa/COSINE_1.00/

The output folder will contain:

• logs: stores the whole log files.
• extraction_cache.pkl: stores the extraction result (e.g., knowledge units and entities).
• graph_cache.json: stores the graph relevant data.
• selection_cache.pkl: stores the hybrid selection result.
• milvus.db: stores the embeddings of three kinds of nodes.
• result.json: stores the generation result.
• retrieval_results.json: stores the retrieval result

Please cite and 🌟 us if you find it helpful!

@misc{su2025clueragaccuratecostefficientgraphbased,
      title={Clue-RAG: Towards Accurate and Cost-Efficient Graph-based RAG via Multi-Partite Graph and Query-Driven Iterative Retrieval}, 
      author={Yaodong Su and Yixiang Fang and Yingli Zhou and Quanqing Xu and Chuanhui Yang},
      year={2025},
      eprint={2507.08445},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2507.08445}, 
}