🤗 HiRAG: Retrieval-Augmented Generation with Hierarchical Knowledge

• This is the repo for the paper HiRAG: Retrieval-Augmented Generation with Hierarchical Knowledge.
• Accepted to EMNLP 2025 Findings!🎉
• Re-indexing the knowledge base is too costly🤯? Want to refine your knowledge base at test time? See our new work DeepRefine!

Model Pipeline

Install

# remember clone this repo first
cd HiRAG
pip install -e .

Quick Start

You can just utilize the following code to perform a query with HiRAG.

graph_func = HiRAG(
    working_dir="./your_work_dir",
    enable_llm_cache=True,
    enable_hierachical_mode=True, 
    embedding_batch_num=6,
    embedding_func_max_async=8, # according to your machine
    enable_naive_rag=True
    )
# indexing
with open("path_to_your_context.txt", "r") as f:
    graph_func.insert(f.read())
# retrieval & generation
print("Perform hi search:")
print(graph_func.query("The question you want to ask?", param=QueryParam(mode="hi")))

Or if you want to employ HiRAG with DeepSeek, ChatGLM, or other third-party retrieval api, here are the examples in ./hi_Search_deepseek.py, ./hi_Search_glm.py, and ./hi_Search_openai.py. The API keys and the LLM configurations can be set at ./config.yaml.

Evaluation

We take the procedure in Mix dataset as an example.

cd ./HiRAG/eval

1. Extract context from original QA datasets.

python extract_context.py -i ./datasets/mix -o ./datasets/mix

2. Insert context to Graph Database.

python insert_context_deepseek.py

Note that the script insert_context_deepseek.py is for the setting of generation with DeepSeek-v3 api, you can replace that with insert_context_openai.py or insert_context_glm.py.

3. Test with different versions of HiRAG.

# there are different retrieval options
# If you want to employ HiRAG approach, just run:
python test_deepseek.py -d mix -m hi
# If you want to employ naive RAG approach, just run:
python test_deepseek.py -d mix -m naive
# If you want to employ HiRAG approach w/o bridge, just run:
python test_deepseek.py -d mix -m hi_nobridge
# If you want to employ HiRAG approach with retrieving only local knowledge, just run:
python test_deepseek.py -d mix -m hi_local
# If you want to employ HiRAG approach with retrieving only global knowledge, just run:
python test_deepseek.py -d mix -m hi_global
# If you want to employ HiRAG approach with retrieving only bridge knowledge, just run:
python test_deepseek.py -d mix -m hi_bridge

Note that the dataset mix can be replaced to any other datasets in Hugging Face link. And the script test_deepseek.py is for the setting of generation with DeepSeek-v3 api, you can replace that with test_openai.py or test_glm.py.

4. Evaluate the generated answers.

First step, request for evaluations.

python batch_eval.py -m request -api openai
python batch_eval.py -m request -api deepseek

Second step, get the results.

python batch_eval.py -m result -api openai
python batch_eval.py -m result -api deepseek

Results

Compare with Naive RAG:

With the config output_file set as f"./datasets/{DATASET}/{DATASET}_eval_hi_naive.jsonl", just run the command:

python batch_eval.py -m result -api openai

Dataset	Dimension	NaiveRAG %	HiRAG %
Mix
	Comprehensiveness	16.6	83.4
	Empowerment	11.6	88.4
	Diversity	12.7	87.3
	Overall	12.4	87.6
CS
	Comprehensiveness	30.0	70.0
	Empowerment	29.0	71.0
	Diversity	14.5	85.5
	Overall	26.5	73.5
Legal
	Comprehensiveness	32.5	67.5
	Empowerment	25.0	75.0
	Diversity	22.0	78.0
	Overall	22.5

Compare with GraphRAG:

With the config output_file set as f"./datasets/{DATASET}/{DATASET}_eval_hi_graphrag.jsonl", just run the command:

python batch_eval.py -m result -api openai

Dataset	Dimension	GraphRAG %	HiRAG %
Mix
	Comprehensiveness	42.1	57.9
	Empowerment	35.1	64.9
	Diversity	40.5	59.5
	Overall	35.9	64.1
CS
	Comprehensiveness	40.5	59.5
	Empowerment	38.5	61.5
	Diversity	30.5	69.5
	Overall	36.0	64.0
Legal
	Comprehensiveness	48.5	51.5
	Empowerment	43.5	56.5
	Diversity	47.0	53.0
	Overall	45.5

Compare with LightRAG:

With the config output_file set as f"./datasets/{DATASET}/{DATASET}_eval_hi_lightrag.jsonl", just run the command:

python batch_eval.py -m result -api openai

Dataset	Dimension	LightRAG %	HiRAG %
Mix
	Comprehensiveness	36.8	63.2
	Empowerment	34.9	65.1
	Diversity	34.1	65.9
	Overall	34.1	65.9
CS
	Comprehensiveness	44.5	55.5
	Empowerment	41.5	58.5
	Diversity	33.0	67.0
	Overall	41.0	59.0
Legal
	Comprehensiveness	49.0	51.0
	Empowerment	43.5	56.5
	Diversity	63.0	37.0
	Overall	48.0

Compare with FastGraphRAG:

With the config output_file set as f"./datasets/{DATASET}/{DATASET}_eval_hi_fastgraphrag.jsonl", just run the command:

python batch_eval.py -m result -api openai

Dataset	Dimension	FastGraphRAG %	HiRAG %
Mix
	Comprehensiveness	0.8	99.2
	Empowerment	0.8	99.2
	Diversity	0.8	99.2
	Overall	0.8	99.2
CS
	Comprehensiveness	0.0	100.0
	Empowerment	0.0	100.0
	Diversity	0.5	99.5
	Overall	0.0	100.0
Legal
	Comprehensiveness	1.0	99.0
	Empowerment	0.0	100.0
	Diversity	1.5	98.5
	Overall	0.0

Compare with KAG:

With the config output_file set as f"./datasets/{DATASET}/{DATASET}_eval_hi_kag.jsonl", just run the command:

python batch_eval.py -m result -api openai

Dataset	Dimension	KAG %	HiRAG %
Mix
	Comprehensiveness	2.3	97.7
	Empowerment	3.5	96.5
	Diversity	3.8	96.2
	Overall	2.3	97.7
CS
	Comprehensiveness	1.0	99.0
	Empowerment	4.5	95.5
	Diversity	5.0	95.0
	Overall	1.5	98.5
Legal
	Comprehensiveness	16.5	83.5
	Empowerment	9.0	91.0
	Diversity	11.0	89.0
	Overall	8.5

Acknowledgement

We gratefully acknowledge the use of the following open-source projects in our work:

• nano-graphrag: a simple, easy-to-hack GraphRAG implementation

• RAPTOR: a novel approach to retrieval-augmented language models by constructing a recursive tree structure from documents.

💡 Cite Us

@article{huang2025retrieval,
  title={Retrieval-Augmented Generation with Hierarchical Knowledge},
  author={Huang, Haoyu and Huang, Yongfeng and Yang, Junjie and Pan, Zhenyu and Chen, Yongqiang and Ma, Kaili and Chen, Hongzhi and Cheng, James},
  journal={arXiv preprint arXiv:2503.10150},
  year={2025}
}

🌟 Star History