VecTree-RAG: Agentic Framework for Scientific QA

VecTree-RAG is an agentic retrieval-augmented generation framework that separates paper identification from evidence localization using vector search and tree traversal. It addresses the limitation of conventional RAG systems that flatten document structure and ignore methodological context. Vector search ranks document and section representations across the corpus, while reasoning-guided traversal of source-verified section trees finds evidence within shortlisted papers. Full text is stored in a page store and revealed only after structural localization. The framework was evaluated on 300 QASPER questions, 54 LitQA2 questions, and 49 multi-document queries, demonstrating improved efficiency and accuracy.