VDGR-RAG Unifies Enterprise Knowledge Reasoning for QA.

Retrieval-Augmented Generation (RAG) is vital for enterprise knowledge question answering, particularly in sectors with extensive and complex documentation, such as telecommunications. However, current RAG methods often fail to fully integrate diverse retrieval strengths, leading to inaccurate domain routing and poor utilization of hierarchical document structures, which limits their reasoning capabilities. To overcome these challenges, VDGR-RAG (Vectors, Directories, Graphs, and Reflection) proposes a unified framework. This agentic GraphRAG system first constructs a Hierarchical Heterogeneous Knowledge Graph (H²KG) from document chunks, preserving both hierarchical directory structures and semantic relationships. This H²KG serves as the foundation for its advanced retrieval mechanisms. VDGR-RAG employs a suite of atomic tools that can be flexibly combined to navigate the H²KG. These include a directory-enhanced routing tool using tables of contents for domain-specific query routing, a multi-route retrieval tool combining vector, TOC-based, and graph searches, a directory backtracking tool to correct localization biases, and a dynamic reflection tool for iterative retrieval planning. Extensive experiments on enterprise product documents demonstrate VDGR-RAG's superior performance in knowledge retrieval recall and QA accuracy compared to existing RAG baselines.