TSA-RAG: Structure-Preserving Retrieval-Augmented Generation via Tiered Document Graph
Abstract
AbstractRetrieval-Augmented Generation (RAG) effectively mitigates hallucinations and knowledge obsolescence in Large Language Models (LLMs). However, prevailing RAG systems typically retrieve evidence from discrete chunks or through decoupled structures, making it difficult to jointly preserve cross-document navigability and within-document coherence on complex tasks. To address this challenge, we propose TSA-RAG (Tiered-Structure-Aware RAG), a unified framework that recasts retrieval as structure-preserving navigation over a Tiered Document Graph (TDG). The TDG offers a joint retrieval space that explicitly couples relational topology with hierarchical skeletons, enabling seamless traversal across granularities. Building on this representation, TSA-RAG navigates evidence through a progressive Expand-Route-Retrieve paradigm, which expands query-focused subgraphs under path-consistency constraints, routes the query to structural entry points within relevant documents, and retrieves subtree-structured evidence preserving discourse continuity. Extensive experiments on seven QA datasets demonstrate that TSA-RAG consistently outperforms strong baselines while yielding more compact retrieval contexts. Quantitative analyses further confirm the scalability and robustness of our framework. Code is available at https://github.com/ZKYQUQ/TSA-RAG.