Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Graph-Anchored Knowledge Indexing for Retrieval-Augmented Generation

Zhenghao Liu, Mingyan Wu, Xinze Li, Yukun Yan, Shuo Wang, Cheng Yang, Minghe Yu, Zheni Zeng, Maosong Sun

TL;DR

GraphAnchor addresses retrieval-induced noise in RAG by maintaining an evolving knowledge graph that anchors salient entities and relations during iterative retrieval. The graph acts as a context-aware index, guiding reasoning, next-query generation, and final answer synthesis, while expanding its coverage of linked concepts over time. Across four multi-hop QA benchmarks, GraphAnchor yields consistent improvements over Vanilla RAG and other iterative baselines, with ablations confirming the critical roles of both iterative retrieval and graph-based indexing. The approach also demonstrates that the evolving graph modulates LLM attention to focus on evidence-grounded graph anchors, aiding coherent integration of distributed information.

Abstract

Retrieval-Augmented Generation (RAG) has emerged as a dominant paradigm for mitigating hallucinations in Large Language Models (LLMs) by incorporating external knowledge. Nevertheless, effectively integrating and interpreting key evidence scattered across noisy documents remains a critical challenge for existing RAG systems. In this paper, we propose GraphAnchor, a novel Graph-Anchored Knowledge Indexing approach that reconceptualizes graph structures from static knowledge representations into active, evolving knowledge indices. GraphAnchor incrementally updates a graph during iterative retrieval to anchor salient entities and relations, yielding a structured index that guides the LLM in evaluating knowledge sufficiency and formulating subsequent subqueries. The final answer is generated by jointly leveraging all retrieved documents and the final evolved graph. Experiments on four multi-hop question answering benchmarks demonstrate the effectiveness of GraphAnchor, and reveal that GraphAnchor modulates the LLM's attention to more effectively associate key information distributed in retrieved documents. All code and data are available at https://github.com/NEUIR/GraphAnchor.

Graph-Anchored Knowledge Indexing for Retrieval-Augmented Generation

TL;DR

GraphAnchor addresses retrieval-induced noise in RAG by maintaining an evolving knowledge graph that anchors salient entities and relations during iterative retrieval. The graph acts as a context-aware index, guiding reasoning, next-query generation, and final answer synthesis, while expanding its coverage of linked concepts over time. Across four multi-hop QA benchmarks, GraphAnchor yields consistent improvements over Vanilla RAG and other iterative baselines, with ablations confirming the critical roles of both iterative retrieval and graph-based indexing. The approach also demonstrates that the evolving graph modulates LLM attention to focus on evidence-grounded graph anchors, aiding coherent integration of distributed information.

Abstract

Retrieval-Augmented Generation (RAG) has emerged as a dominant paradigm for mitigating hallucinations in Large Language Models (LLMs) by incorporating external knowledge. Nevertheless, effectively integrating and interpreting key evidence scattered across noisy documents remains a critical challenge for existing RAG systems. In this paper, we propose GraphAnchor, a novel Graph-Anchored Knowledge Indexing approach that reconceptualizes graph structures from static knowledge representations into active, evolving knowledge indices. GraphAnchor incrementally updates a graph during iterative retrieval to anchor salient entities and relations, yielding a structured index that guides the LLM in evaluating knowledge sufficiency and formulating subsequent subqueries. The final answer is generated by jointly leveraging all retrieved documents and the final evolved graph. Experiments on four multi-hop question answering benchmarks demonstrate the effectiveness of GraphAnchor, and reveal that GraphAnchor modulates the LLM's attention to more effectively associate key information distributed in retrieved documents. All code and data are available at https://github.com/NEUIR/GraphAnchor.
Paper Structure (20 sections, 8 equations, 14 figures, 5 tables)

This paper contains 20 sections, 8 equations, 14 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Graph-based Knowledge Anchoring for Iterative RAG
  5. Stepwise Graph Evolution for Context-Aware Knowledge Indexing
  6. Experimental Methodology
  7. Evaluation Result
  8. Overall Performance
  9. Ablation Study
  10. Effectiveness of Graphs in GraphAnchor During Iterative Retrieval
  11. Attention Visualization by Incorporating Knowledge Anchoring
  12. Conclusion
  13. Appendix
  14. License
  15. Additional Experimental Details
  16. ...and 5 more sections

Figures (14)

  • Figure 1: Illustration of GraphAnchor. The graph is progressively updated for knowledge indexing.
  • Figure 2: Illustration of GraphAnchor.
  • Figure 3: Knowledge Anchoring Effectiveness across Different Retrieval Steps.
  • Figure 4: Effectiveness of the Constructed Knowledge Graph for Retrieval. Figures \ref{['fig:passage_hitrate']} and \ref{['fig:overlap']} report the retrieval hit rate and document overlap, respectively, to evaluate accuracy and diversity.
  • Figure 5: Attention Distribution when Answering Queries with Documents and Graphs in GraphAnchor.
  • ...and 9 more figures