PT-RAG: Structure-Fidelity Retrieval-Augmented Generation for Academic Papers

ABSTRACT

PT-RAG proposes a structure-aware retrieval framework designed specifically for long academic documents. Traditional RAG pipelines flatten documents into chunks, which destroys the natural hierarchical structure of papers. PT-RAG instead constructs a hierarchical PaperTree index that preserves section relationships and document structure. Retrieval then follows query-aligned root-to-leaf paths through the document tree. This design reduces context fragmentation and improves evidence alignment during generation.