Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

T-Retriever: Tree-based Hierarchical Retrieval Augmented Generation for Textual Graphs

Chunyu Wei, Huaiyu Qin, Siyuan He, Yunhai Wang, Yueguo Chen

TL;DR

T-Retriever rethinks graph-based RAG by reframing retrieval over textual attributed graphs as tree-based retrieval guided by Semantic-Structural Entropy (S²-Entropy). It couples Adaptive Compression Encoding, a top-down partitioning strategy inspired by Shannon-Fano coding, with KDE-based semantic density to form a hierarchical Encoding Tree that preserves both structure and semantics. The framework embeds and indexes summaries at each node, retrieves relevant subgraphs via an ANN-enabled multi-level index, and uses a GNN-augmented LLM prompting pipeline to generate answers, achieving state-of-the-art results on multiple graph reasoning benchmarks while significantly reducing online context size. Empirically, T-Retriever shows consistent gains, especially on larger graphs, and demonstrates a favorable offline-online efficiency balance, with no requirement for additional parameter updates during deployment.

Abstract

Retrieval-Augmented Generation (RAG) has significantly enhanced Large Language Models' ability to access external knowledge, yet current graph-based RAG approaches face two critical limitations in managing hierarchical information: they impose rigid layer-specific compression quotas that damage local graph structures, and they prioritize topological structure while neglecting semantic content. We introduce T-Retriever, a novel framework that reformulates attributed graph retrieval as tree-based retrieval using a semantic and structure-guided encoding tree. Our approach features two key innovations: (1) Adaptive Compression Encoding, which replaces artificial compression quotas with a global optimization strategy that preserves the graph's natural hierarchical organization, and (2) Semantic-Structural Entropy ($S^2$-Entropy), which jointly optimizes for both structural cohesion and semantic consistency when creating hierarchical partitions. Experiments across diverse graph reasoning benchmarks demonstrate that T-Retriever significantly outperforms state-of-the-art RAG methods, providing more coherent and contextually relevant responses to complex queries.

T-Retriever: Tree-based Hierarchical Retrieval Augmented Generation for Textual Graphs

TL;DR

T-Retriever rethinks graph-based RAG by reframing retrieval over textual attributed graphs as tree-based retrieval guided by Semantic-Structural Entropy (S²-Entropy). It couples Adaptive Compression Encoding, a top-down partitioning strategy inspired by Shannon-Fano coding, with KDE-based semantic density to form a hierarchical Encoding Tree that preserves both structure and semantics. The framework embeds and indexes summaries at each node, retrieves relevant subgraphs via an ANN-enabled multi-level index, and uses a GNN-augmented LLM prompting pipeline to generate answers, achieving state-of-the-art results on multiple graph reasoning benchmarks while significantly reducing online context size. Empirically, T-Retriever shows consistent gains, especially on larger graphs, and demonstrates a favorable offline-online efficiency balance, with no requirement for additional parameter updates during deployment.

Abstract

Retrieval-Augmented Generation (RAG) has significantly enhanced Large Language Models' ability to access external knowledge, yet current graph-based RAG approaches face two critical limitations in managing hierarchical information: they impose rigid layer-specific compression quotas that damage local graph structures, and they prioritize topological structure while neglecting semantic content. We introduce T-Retriever, a novel framework that reformulates attributed graph retrieval as tree-based retrieval using a semantic and structure-guided encoding tree. Our approach features two key innovations: (1) Adaptive Compression Encoding, which replaces artificial compression quotas with a global optimization strategy that preserves the graph's natural hierarchical organization, and (2) Semantic-Structural Entropy (-Entropy), which jointly optimizes for both structural cohesion and semantic consistency when creating hierarchical partitions. Experiments across diverse graph reasoning benchmarks demonstrate that T-Retriever significantly outperforms state-of-the-art RAG methods, providing more coherent and contextually relevant responses to complex queries.
Paper Structure (24 sections, 2 theorems, 14 equations, 4 figures, 4 tables)

This paper contains 24 sections, 2 theorems, 14 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Graph-based Retrieval for Large Language Models.
  4. Hierarchical Indexing for RAG
  5. Structural Entropy and Our Novelty
  6. Preliminaries
  7. Textual Attributed Graphs.
  8. Retrieval-Augmented Generation (RAG) for Graphs.
  9. Encoding Tree.
  10. Structural Entropy.
  11. Methodology
  12. Semantic-Structural Entropy
  13. Adaptive Compression Encoding
  14. Indexing with Encoding Tree
  15. Retrieving from Encoding Tree
  16. ...and 9 more sections

Key Result

Proposition 1

For any encoding tree $\mathcal{T}$ and nodes $\alpha, \beta \in \mathcal{T}$ where $\alpha$ is an ancestor of $\beta$ with height difference >1, the Regulate operation preserves S²-Entropy: $H_{S^2}^{\mathcal{T}}(\mathcal{G}) = H_{S^2}^{\mathcal{T}_{\text{REGULATE}(\alpha, \beta)}}(\mathcal{G})$. P

Figures (4)

  • Figure 1: Illustration of T-Retriever. Hierarchical organization of attributed graph knowledge enabling effective multi-resolution context retrieval for question answering.
  • Figure 2: The T-Retriever framework pipeline: (1) Encoding Tree Construction optimizes S²-Entropy (combining structural and semantic information) through partition, prune, and regulate operations; (2) Indexing generates and embeds LLM-based summaries for tree nodes; (3) Tree Retrieval finds relevant nodes, extracts subgraphs, and generates responses using GNN-enhanced LLM prompting.
  • Figure 3: Sensitivity analysis of key hyperparameters. (a-c) Impact of encoding tree layers $L$ and retrieved subgraphs $k$ across different datasets. (d) Impact of the entropy weighting factor $\lambda$ on WebQSP.
  • Figure 4: Case study visualization from BookGraphs.

Theorems & Definitions (5)

  • Definition 1: Partition Operation
  • Definition 2: Prune Operation
  • Definition 3: Regulate Operation
  • Proposition 1
  • Proposition 2: Catalytic Effect