LeanRAG: Knowledge-Graph-Based Generation with Semantic Aggregation and Hierarchical Retrieval

🎉 This paper has been accepted by AAAI-26! 🎉

LeanRAG is an efficient, open-source framework for Retrieval-Augmented Generation, leveraging knowledge graph structures with semantic aggregation and hierarchical retrieval to generate context-aware, concise, and high-fidelity responses.

✨ Features

• Semantic Aggregation: Clusters entities into semantically coherent summaries and constructs explicit relations to form a navigable aggregation-level knowledge network.
• Hierarchical, Structure-Guided Retrieval: Initiates retrieval from fine-grained entities and traverses up the knowledge graph to gather rich, highly relevant evidence efficiently.
• Reduced Redundancy: Optimizes retrieval paths to significantly reduce redundant information—LeanRAG achieves ~46% lower retrieval redundancy compared to flat retrieval baselines (based on benchmark evaluations).
• Benchmark Performance: Demonstrates superior performance across multiple QA benchmarks with improved response quality and retrieval efficiency.

🏛️ Architecture Overview

LeanRAG’s processing pipeline follows these core stages:

1. Semantic Aggregation

2. Group low-level entities into clusters; generate summary nodes and build adjacency relations among them for efficient navigation.

3. Knowledge Graph Construction

4. Construct a multi-layer graph where nodes represent entities and aggregated summaries, with explicit inter-node relations for graph-based traversal.

5. Query Processing & Hierarchical Retrieval

6. Anchor queries at the most relevant detailed entities ("bottom-up"), then traverse upward through the semantic aggregation graph to collect evidence spans.

7. Redundancy-Aware Synthesis

8. Streamline retrieval paths and avoid overlapping content, ensuring concise evidence aggregation before generating responses.

9. Generation

10. Use retrieved, well-structured evidence as input to an LLM to produce coherent, accurate, and contextually grounded answers.

🚀 Getting Started

Prerequisites

• Python 3.10+
• Conda for environment management

Installation

1. Clone the repository: bash git clone https://github.com/RaZzzyz/LeanRAG.git cd LeanRAG

2. Create a virtual environment: bash conda install -n leanrag python=3.11 conda activate leanrag

3. Install the required dependencies: bash pip install -r requirements.txt

💻 Usage Workflow

Here’s a typical pipeline flow:

Step 1: Document Chunking

In file_chunk.py, split the document into chunks:

• Chunk size: 1024
• Sliding step: 128 (i.e., use a sliding window with step 128)

Each dictionary in the resulting chunk file contains two attributes:

• hash_code: hash calculated from the text content for traceability
• text: the chunk text content

---

Step 2: Extract Triples and Entity Descriptions

Two knowledge graph extraction methods are currently provided:

Method 1: CommonKG

Based on Wikipedia entities. First, define a head entity list, then extract triples from the document.

Usage:

1. Edit the configuration file:
   CommonKG/config/create_kg_conf_test.yaml
   Fill in the model's url and name, and the path to the chunk file.
2. Run extraction:
   bash python CommonKG/create_kg.py The extraction result will be saved in output_dir.
3. Process 6-tuples with descriptions: python python CommonKG/deal_triple.py Outputs include:
   • entity.jsonl
   • relation.jsonl

Method 2: GraphRAG

Relies on LLM capability to perform few-shot extraction with given examples in the prompt.

Usage:

1. Edit GraphExtraction/chunk.py to fill in url and model. The chunk_file is the same as in CommonKG, generated from Step 1.
2. Deduplicate extraction results: python python GraphExtraction/deal_triple.py Outputs include:
   • entity.jsonl
   • relation.jsonl

Step 3: Build the Graph

python build_graph.py

• Cluster extracted entity and relation descriptions and generate relationships.

• Construct a tree-structured knowledge graph, supporting retrieval and Q&A.

Step 3: Retrieval

python query_graph.py

1. Select the correct chunks_file.

2. Query the graph for Top-K entities based on query.

3. Generate paths between nodes according to the tree structure.

4. Return same-level relationships and aggregated entity information along the paths to the LLM for final answer generation.

📊 Results & Benchmarks

On four challenging QA benchmarks spanning diverse domains, LeanRAG consistently delivers:

Score

Mix

Metric	LeanRAG	HiRAG	Naive	GraphRAG	LightRAG	FastGraphRAG	KAG
Comprehensiveness	8.89±0.01	8.72±0.02	8.20±0.01	8.52±0.01	8.19±0.02	6.56±0.02	7.90±0.03
Empowerment	8.16±0.02	7.86±0.03	7.52±0.03	7.73±0.02	7.56±0.03	5.82±0.03	7.41±0.04
Diversity	7.73±0.01	7.21±0.02	6.65±0.03	7.04±0.02	6.69±0.04	4.88±0.03	6.42±0.04
Overall	8.59±0.01	8.08±0.02	7.47±0.02	7.87±0.01	7.61±0.04	5.76±0.02	7.25±0.03

---

CS

Metric	LeanRAG	HiRAG	Naive	GraphRAG	LightRAG	FastGraphRAG	KAG
Comprehensiveness	8.92±0.01	8.92±0.01	8.94±0.01	8.55±0.02	8.76±0.02	6.79±0.01	8.22±0.02
Empowerment	8.68±0.02	8.66±0.02	8.69±0.04	8.28±0.04	8.50±0.04	6.67±0.04	8.52±0.05
Diversity	7.87±0.02	7.84±0.02	7.79±0.02	7.42±0.02	7.63±0.04	5.45±0.04	7.03±0.02
Overall	8.82±0.02	8.77±0.02	8.77±0.03	8.37±0.04	8.59±0.04	6.31±0.03	7.99±0.03

---

Legal

Metric	LeanRAG	HiRAG	Naive	GraphRAG	LightRAG	FastGraphRAG	KAG
Comprehensiveness	8.88±0.02	8.68±0.02	8.85±0.01	8.95±0.01	8.24±0.02	3.87±0.02	8.41±0.02
Empowerment	8.42±0.03	8.18±0.06	8.28±0.03	8.33±0.02	7.83±0.05	3.53±0.03	8.20±0.03
Diversity	7.49±0.03	7.00±0.03	7.10±0.04	7.47±0.03	6.87±0.01	2.87±0.02	6.71±0.01
Overall	8.49±0.04	8.00±0.04	8.21±0.03	8.44±0.01	7.74±0.03	3.43±0.02	7.83±0.03

---

Agriculture

Metric	LeanRAG	HiRAG	Naive	GraphRAG	LightRAG	FastGraphRAG	KAG
Comprehensiveness	8.94±0.06	8.99±0.00	8.85±0.01	8.97±0.01	8.71±0.01	3.28±0.01	8.22±0.01
Empowerment	8.66±0.02	8.52±0.02	8.51±0.03	8.52±0.02	8.23±0.02	3.29±0.05	8.33±0.06
Diversity	8.06±0.03	7.98±0.02	7.76±0.06	7.95±0.02	7.68±0.03	3.01±0.03	7.07±0.02
Overall	8.87±0.02	8.87±0.03	8.69±0.03	8.85±0.01	8.56±0.02	3.17±0.02	7.95±0.03
### Winrate

NaiveRAG vs LeanRAG

Metric	Mix (NaiveRAG)	Mix (LeanRAG)	CS (NaiveRAG)	CS (LeanRAG)	Legal (NaiveRAG)	Legal (LeanRAG)	Agriculture (NaiveRAG)	Agriculture (LeanRAG)
Comprehensiveness	11.9%	88.1%	41.0%	59.0%	30.0%	70.0%	37.7%	62.3%
Empowerment	1.5%	98.5%	40.5%	59.5%	24.5%	75.5%	19.8%	80.2%
Diversity	3.1%	96.9%	28.0%	72.0%	9.0%	91.0%	10.0%	90.0%
Overall	2.7%	97.3%	39.5%	60.5%	23.5%	76.5%	19.3%	80.7%

GraphRAG vs LeanRAG

Metric	Mix (GraphRAG)	Mix (LeanRAG)	CS (GraphRAG)	CS (LeanRAG)	Legal (GraphRAG)	Legal (LeanRAG)	Agriculture (GraphRAG)	Agriculture (LeanRAG)
Comprehensiveness	35.0%	65.0%	41.0%	59.0%	49.0%	51.0%	45.5%	54.5%
Empowerment	20.0%	80.0%	33.5%	66.5%	44.0%	56.0%	27.0%	73.0%
Diversity	16.5%	83.5%	34.0%	66.0%	44.0%	56.0%	22.0%	78.0%
Overall	21.9%	78.1%	37.5%	62.5%	47.0%	53.0%	28.5%	71.5%

LightRAG vs LeanRAG

Metric	Mix (LightRAG)	Mix (LeanRAG)	CS (LightRAG)	CS (LeanRAG)	Legal (LightRAG)	Legal (LeanRAG)	Agriculture (LightRAG)	Agriculture (LeanRAG)
Comprehensiveness	28.8%	71.2%	44.5%	55.5%	25.0%	75.0%	38.0%	62.0%
Empowerment	16.5%	83.5%	35.5%	64.5%	12.0%	88.0%	17.0%	83.0%
Diversity	13.1%	86.9%	34.0%	66.0%	40.5%	59.5%	16.5%	83.5%
Overall	18.8%	81.2%	38.5%	61.5%	21.0%	79.0%	18.5%	81.5%

FastGraphRAG vs LeanRAG

Metric	Mix (FastGraphRAG)	Mix (LeanRAG)	CS (FastGraphRAG)	CS (LeanRAG)	Legal (FastGraphRAG)	Legal (LeanRAG)	Agriculture (FastGraphRAG)	Agriculture (LeanRAG)
Comprehensiveness	0.0%	100.0%	0.5%	99.5%	1.0%	99.0%	0.5%	99.5%
Empowerment	0.0%	100.0%	0.0%	100.0%	0.5%	99.5%	0.0%	100.0%
Diversity	0.0%	100.0%	0.8%	99.2%	2.5%	97.5%	0.0%	100.0%
Overall	0.0%	100.0%	0.0%	100.0%	4.5%	95.5%	0.0%	100.0%

KAG vs LeanRAG

Metric	Mix (KAG)	Mix (LeanRAG)	CS (KAG)	CS (LeanRAG)	Legal (KAG)	Legal (LeanRAG)	Agriculture (KAG)	Agriculture (LeanRAG)
Comprehensiveness	1.5%	98.5%	5.0%	95.0%	5.0%	95.0%	2.5%	97.5%
Empowerment	1.9%	98.1%	3.0%	97.0%	4.5%	95.5%	2.5%	97.5%
Diversity	1.2%	98.8%	4.0%	96.0%	2.5%	97.5%	1.0%	99.0%
Overall	1.2%	98.8%	3.5%	96.5%	4.5%	95.5%	1.0%	99.0%

HiRAG vs LeanRAG

Metric	Mix (HiRAG)	Mix (LeanRAG)	CS (HiRAG)	CS (LeanRAG)	Legal (HiRAG)	Legal (LeanRAG)	Agriculture (HiRAG)	Agriculture (LeanRAG)
Comprehensiveness	43.8%	56.2%	46.5%	53.5%	29.5%	70.5%	49.5%	50.5%
Empowerment	26.5%	73.5%	43.5%	56.5%	16.5%	83.5%	26.5%	73.5%
Diversity	20.4%	79.6%	44.5%	55.5%	23.5%