VecTree-RAG beats dense RAG with 0.925 accuracy on scientific QA
New framework combines vector and tree search for 6x better evidence precision.
VecTree-RAG is a new retrieval-augmented generation framework designed for scientific question answering. Traditional RAG flattens documents into fixed-length passages, missing the structural hierarchy of research papers. VecTree-RAG solves this with a two-stage process: first, vector search ranks compact representations of documents and sections across the entire corpus; second, a reasoning-guided traversal of source-verified section trees within shortlisted papers pinpoints exact evidence pages. The full text is only exposed progressively after structural localization, minimizing irrelevant content.
Evaluated on three benchmarks—300 QASPER, 54 LitQA2, and 49 MOSAIC questions—VecTree-RAG achieved top scores: 0.800 LLM-judge correctness on QASPER, 0.925 accuracy on LitQA2, and a 0.547 composite on MOSAIC. Crucially, evidence-page precision was 0.274, a 4–6x improvement over baselines like Dense RAG and RAPTOR. Ablation studies confirmed that both the vector and tree components are essential. While multi-turn inference is pricier than single-call retrieval, the framework offers structure-aware, traceable evidence extraction—ideal for researchers and analysts.
- VecTree-RAG combines vector search for corpus retrieval and tree traversal for evidence localization, achieving 0.274 evidence-page precision vs. 0.046–0.071 for baselines.
- Outperformed Dense RAG, RAPTOR, and Search-o1 on QASPER (0.800), LitQA2 (0.925), and MOSAIC (0.547).
- Full text is only loaded after structural localization, reducing wasteful inference tokens and focusing LLM attention on relevant sections.
Why It Matters
Scientific literature QA becomes far more accurate and efficient by preserving document structure instead of flattening it.