[2E1-GS-5b] Agents

9:45 AM - 10:00 AM JST(12:45 AM - 1:00 AM UTC)

[2E1-GS-5b-04] A Study on Improving the Accuracy of Agentic RAG Based on Directory Hierarchical Structures

〇Shogo Namba¹, Makiko Naemura², Kenzo Kurotsuchi², Ryo Tanabe¹, Yuxin Liang¹, Kodai Takeda¹ (1. Hitachi Ltd.,, 2. Hitachi Industrial Equipment Systems Co., Ltd.)

Keywords:

Agent,RAG,Agentic RAG

Retrieval-Augmented Generation (RAG) is a standard method for Large Language Models (LLMs) to supplement knowledge and generate factual responses. However, traditional vector search faces a critical challenge: fragmenting documents into "chunks" causes an irreversible loss of structural context and conceptual knowledge structures (Ontology). Consequently, conventional methods struggle to account for hierarchical dependencies in complex knowledge bases, leading to reduced retrieval accuracy.This study aims to improve retrieval and generation precision in complex environments by integrating the semantic information inherent in document storage structures into the search process. We propose Agentic RAG, which treats existing directory hierarchies as explicit Semantic Hierarchies. In this approach, LLM-based autonomous agents actively explore and navigate these structures to extract relevant information.Experimental results using large-scale repositories demonstrate that the proposed method outperforms traditional vector search in both retrieval accuracy and contextual relevance. These benefits are especially pronounced in complex data environments where similar topics are dispersed across multiple hierarchical layers.