Graph-Anchored Knowledge Indexing for Retrieval-Augmented Generation

TL;DR

GraphAnchor introduces an innovative graph-based indexing method that dynamically updates during retrieval, enhancing the integration of scattered evidence for improved large language model responses in multi-hop question answering.

Contribution

It presents a novel, evolving graph-anchored indexing technique that improves evidence integration and reasoning in retrieval-augmented generation systems.

Findings

Outperforms baseline methods on four multi-hop QA benchmarks.

Enhances LLM's ability to associate key information across documents.

Modulates LLM attention for better evidence utilization.

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.