GraphRAG를 위한 에이전트 기반 서브그래프 탐색 기법

초록

To improve the multi-hop question-answering performance of GraphRAG, this paper proposes an agent-based subgraph explorationtechnique in which LLM agents autonomously navigate a knowledge graph to retrieve query-relevant information.. Specifically, we designeda dual-agent architecture consisting of an Answer Agent, which generates responses or identifies information gaps based on the currentsubgraph, and an Expansion Agent, which selects the most relevant nodes from a pool of candidates to extend the subgraph. Our preliminaryexperiments revealed a significant efficiency bottleneck: providing all candidate nodes to the Expansion Agent without filtering resultedin an average consumption of 21,849 prompt tokens per query. To address this, we introduce a candidate filtering method that scorescandidates by weight-averaging the semantic similarity of node embeddings and the relational similarity of edge embeddings, passingonly the top- candidates to the Expansion Agent. In evaluations using the HotpotQA dataset, the proposed method achieved an F1 scoreof 0.416 and an Exact Match of 0.314, while reducing prompt token consumption by approximately 4.5 times (4,858 vs. 21,849 tokens)compared to the unfiltered baseline (F1 0.414).

키워드