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

Conference: AAAI 2026 arXiv: 2508.10391 Code: GitHub Area: Other Keywords: Retrieval-Augmented Generation, Knowledge Graph, Hierarchical Aggregation, Lowest Common Ancestor, Semantic Network

TL;DR

This paper proposes LeanRAG, a framework that employs a semantic aggregation algorithm to automatically construct explicit relations among summary nodes in a hierarchical knowledge graph, thereby breaking "semantic islands." Combined with a bottom-up retrieval strategy based on the Lowest Common Ancestor (LCA), LeanRAG efficiently navigates the hierarchical structure, achieving state-of-the-art performance on four QA benchmarks while reducing retrieval redundancy by 46%.

Background & Motivation

Retrieval-Augmented Generation (RAG) mitigates hallucination in LLMs by coupling them with external knowledge bases. However, naive RAG's chunk-based retrieval suffers from the "chunking dilemma": fine-grained chunks lose context, while coarse-grained chunks introduce noise. Knowledge-graph-based RAG methods have emerged to address this—GraphRAG organizes documents into community knowledge graphs to preserve local context, and HiRAG further introduces hierarchical structures that cluster entities into multi-level summaries.

Nevertheless, existing hierarchical KG-RAG methods leave two critical challenges unresolved:

Semantic Island Problem: High-level summary nodes lack explicit relational connections among themselves, leaving them mutually isolated. Cross-community reasoning between different concept communities is impossible, turning high-level knowledge into a collection of disconnected "islands."

Structure–Retrieval Decoupling Problem: The retrieval process is topology-agnostic and often degenerates into flat semantic search over all nodes. The carefully constructed hierarchical index is "flattened" during retrieval; rich structural information is used only for post-retrieval context expansion rather than guiding retrieval itself.

The core idea of this paper is to co-design knowledge structure construction and retrieval strategy: the aggregation phase not only clusters entities but also automatically infers inter-summary relations (eliminating semantic islands), while the retrieval phase natively exploits the hierarchical structure through LCA path traversal for precise navigation (eliminating structure–retrieval decoupling).

Method

Overall Architecture

LeanRAG consists of two core innovations: (1) Hierarchical Knowledge Graph Aggregation, which recursively constructs a multi-layer semantic network \(\mathcal{H} = \{\mathcal{G}_0, \mathcal{G}_1, \ldots, \mathcal{G}_k\}\) bottom-up from a base KG \(\mathcal{G}_0\), with each layer containing progressively more abstract entities and explicit inter-entity relations; and (2) LCA-Based Structured Retrieval, which, given a query, first anchors the most relevant fine-grained entities at the base layer, then traverses upward through the hierarchy to the lowest common ancestor, constructing a compact and contextually coherent subgraph as input to the LLM.

Key Designs

1. Recursive Semantic Clustering:

   • Function: Groups entities at each layer by semantic similarity.
   • Mechanism: Entity description texts are encoded into dense vectors \(\mathbf{E}_{i-1} = \{\Phi(d_v) \mid v \in V_{i-1}\}\) using a pretrained embedding model, and a Gaussian Mixture Model (GMM) partitions entities into \(m\) disjoint clusters \(\mathcal{C}_{i-1} = \{C_1, C_2, \ldots, C_m\}\).
   • Design Motivation: GMM is more flexible than hard clustering and can capture soft boundaries between entities. Clustering based on textual descriptions rather than graph structural features ensures semantic coherence.

2. Aggregated Entity Generation:

   • Function: Generates a more abstract summary entity for each cluster.
   • Mechanism: For each cluster \(C_j\), an LLM-driven generation function \(\mathcal{F}_{\text{entity}}\) synthesizes the intra-cluster entities and their relations to produce a new abstract entity and description: \((\alpha_j, d_{\alpha_j}) = \mathcal{F}_{\text{entity}}(C_j, R_{C_j})\). The new entity \(\alpha_j\) becomes the parent node of all entities in the cluster.
   • Design Motivation: Rather than simple clustering, intra-cluster relations are considered when generating summary entities, thereby preserving relational information in semantically meaningful abstractions.

3. Aggregated Relation Generation:

   • Function: Infers and creates new high-level relations between aggregated entities to break semantic islands.
   • Mechanism: For any pair of aggregated entities \((\alpha_j, \alpha_k)\), the cross-cluster relations \(R_{<C_j, C_k>}\) between their underlying clusters \(C_j\) and \(C_k\) are counted. A connectivity strength \(\lambda_{j,k}\) is defined as the number of cross-cluster relations. If \(\lambda_{j,k}\) exceeds a dynamic threshold \(\tau\), an LLM generates a semantic high-level relation description \(r_{<\alpha_j, \alpha_k>} = \mathcal{F}_{\text{rel}}(\alpha_j, \alpha_k, R_{<C_j, C_k>})\); otherwise, the underlying relation texts are directly concatenated.
   • Design Motivation: This is the core innovation distinguishing LeanRAG from all prior methods. Approaches such as HiRAG cluster entities without establishing inter-summary relations, leaving high-level nodes as islands. Explicit relations ensure the entire hierarchical graph remains navigable at every layer.

4. Initial Entity Anchoring:

   • Function: Anchors the user query to the most relevant fine-grained entities at the base layer.
   • Mechanism: Dense retrieval is performed exclusively over the entity set of \(\mathcal{G}_0\) to identify the top-\(n\) semantically most similar "seed entities": \(V_{\text{seed}} = \text{Top-n}_{v \in V_0}(\text{sim}(q, d_v))\).
   • Design Motivation: Starting from the finest granularity ensures retrieval precision and avoids imprecise matching directly against high-level summaries.

5. LCA Path Traversal:

   • Function: Constructs the minimal compact subgraph connecting the seed entities along the hierarchical structure.
   • Mechanism: The lowest common ancestor \(v_{\text{lca}}\) (the common ancestor of minimum depth) of the seed entity set \(V_{\text{seed}}\) in \(\mathcal{H}\) is identified; all entities and relations along the shortest paths from each seed entity to \(v_{\text{lca}}\) are then collected: \(\mathcal{P}_{\text{lca}} = \bigcup_{v \in V_{\text{seed}}} \text{ShortestPath}_\mathcal{H}(v, v_{\text{lca}})\).
   • The final retrieved subgraph \(\mathcal{G}_{\text{ret}}\) contains path entities \(V_{\text{ret}}\), path relations \(R_{\text{lca}}\), and inter-cluster relations among aggregated entities at the same layer \(R_{\text{inter-cluster}}\), along with the original text chunks corresponding to the base entities as supplementary evidence.
   • Design Motivation: Compared to finding all paths on a flat graph (which introduces numerous noisy intermediate nodes), LCA path traversal constructs a "narrative structure" from specifics to shared concepts, yielding retrieval content that is both compact and semantically coherent. The hierarchical nature ensures complete coverage from concrete facts to high-level concepts.

Loss & Training

LeanRAG is a training-free retrieval framework with no loss function design. Key hyperparameters include the GMM cluster count, the threshold \(\tau\) (controlling the strictness of high-level relation generation, adaptively adjusted per layer), and the number of retrieval anchors \(n\).

Key Experimental Results

Main Results (Four QA Datasets, 1–10 Scoring, LLM-as-Judge)

Dataset	Metric	LeanRAG	HiRAG	GraphRAG	NaiveRAG	LightRAG
Mix	Overall	8.59	8.08	7.87	7.47	7.61
CS	Overall	8.82	8.77	8.37	8.77	8.59
Legal	Overall	8.49	8.00	8.44	8.21	7.74
Agriculture	Overall	8.87	8.87	8.85	8.69	8.56
Mix	Diversity	7.73	7.21	7.04	6.65	6.69
Legal	Empowerment	8.42	8.18	8.33	8.28	7.83

Ablation Study

RQ3: Effect of Cross-Cluster Relations (Win Rate With vs. Without Relation Paths)

Dataset	Metric	LeanRAG win	LeanRAG w/o Relation win
Mix	Overall	53.8%	46.2%
CS	Overall	58.5%	41.5%
Legal	Overall	56.5%	43.5%
Agriculture	Overall	58.0%	42.0%
CS	Diversity	66.0%	34.0%

RQ4: Necessity of Original Text Context

Dataset	Metric	LeanRAG	LeanRAG w/o Context
Mix	Overall	8.59	7.93 ↓
CS	Overall	8.82	8.34 ↓
Legal	Overall	8.49	8.00 ↓
Agriculture	Overall	8.87	8.53 ↓

RQ2: Retrieval Redundancy Analysis

LeanRAG's retrieved context is on average 46% fewer tokens than baseline methods, substantially reducing information redundancy while maintaining or improving answer quality.

Key Findings

• LeanRAG achieves optimal or comparable performance on nearly all metrics across four datasets, with the most pronounced advantage on the Diversity dimension, reflecting the effect of cross-cluster relations in broadening information sources.
• Removing cross-cluster relations causes the sharpest drop in Diversity (CS dataset: 66% vs. 34% win rate), validating the core value of relational bridging in breaking semantic islands.
• Removing original text leads to the most notable declines in Comprehensiveness and Empowerment, indicating that the graph structure serves as a "semantic indexing and navigation system" while true informational richness comes from the raw text.
• A 46% reduction in retrieval context without quality degradation confirms that LCA path traversal localizes relevant information more precisely than flat search.
• On the Agriculture dataset, LeanRAG is on par with HiRAG, suggesting that for domain-focused, structurally simple scenarios, the additional relation construction in hierarchical aggregation yields limited marginal benefit.

Highlights & Insights

• The definition of the "semantic island" problem is highly precise and targets the core shortcoming of existing hierarchical KG-RAG methods.
• The aggregation algorithm's principle of "building not only entities but also relations" is simple yet far-reaching—it transforms a fragmented hierarchical tree into a fully connected semantic network.
• LCA path traversal is an elegant retrieval strategy that applies a classical tree-algorithm concept to the RAG retrieval problem.
• The co-design philosophy for retrieval and indexing structures is broadly applicable—the pain point of prior methods lies precisely in their decoupling.
• The experimental design is clear, with four targeted RQs and ablation analyses that quantify the contribution of each framework component.

Limitations & Future Work

• The hierarchical aggregation process requires multiple LLM calls to generate aggregated entities and relation descriptions, resulting in high construction costs, especially for large-scale knowledge bases.
• The selection of GMM cluster count and threshold \(\tau\) relies on validation-set tuning, lacking an automated hyperparameter determination method.
• LCA retrieval assumes an approximately tree-structured hierarchy, which may be less suitable for highly interconnected graph structures.
• Evaluation relies primarily on LLM judgment (DeepSeek-V3), without human evaluation to validate scoring reliability.
• Dynamic knowledge graph update scenarios remain unexplored—when underlying documents change, must the hierarchical structure be fully rebuilt or can it be updated incrementally?
• All experiments use a single LLM (DeepSeek-V3) as the generator, leaving the framework's performance across LLMs of different parameter scales unverified.

Related Work & Insights

• The evolutionary lineage GraphRAG → LightRAG → HiRAG → LeanRAG is clear: from community graphs to dual-layer frameworks to hierarchical clustering to fully connected hierarchical networks.
• RAPTOR's recursive summarization tree provides early inspiration for hierarchical retrieval but lacks inter-entity relation modeling—LeanRAG supplements this with an explicit relational dimension.
• The LCA concept may inspire other hierarchical retrieval scenarios, such as file system retrieval and ontological reasoning.
• The paradigm of "structure-guided retrieval" also holds reference value for other structured knowledge sources, such as code graphs and mathematical knowledge bases.

Rating

• Novelty: ⭐⭐⭐⭐ (The semantic island problem definition is novel and the LCA retrieval strategy is creative, though the core technical components are relatively standard.)
• Experimental Thoroughness: ⭐⭐⭐⭐ (Four datasets, four RQs, and multiple ablations, but human evaluation is absent.)
• Writing Quality: ⭐⭐⭐⭐ (Framework diagrams are clear and problem motivation is well-developed, though some notation definitions could be more concise.)
• Value: ⭐⭐⭐⭐⭐ (RAG is a prominent research area; this paper addresses practical pain points with a practical, open-source framework.)

Related Papers

• [AAAI 2026] CAE: Hierarchical Semantic Alignment for Image Clustering
• [ICLR 2026] Exchangeability of GNN Representations with Applications to Graph Retrieval
• [AAAI 2026] Improved Differentially Private Algorithms for Rank Aggregation
• [AAAI 2026] Variance Computation for Weighted Model Counting with Knowledge Compilation Approach
• [ICLR 2026] Learning Structure-Semantic Evolution Trajectories for Graph Domain Adaptation