Adaptive Query Routing: A Tier-Based Framework for Hybrid Retrieval Across Financial, Legal, and Medical Documents
Research shows no single RAG method wins; a new hybrid system dynamically routes queries for best results.
A new research paper by Afshan Hashmi, "Adaptive Query Routing: A Tier-Based Framework for Hybrid Retrieval," provides a rigorous comparison of three major Retrieval-Augmented Generation (RAG) architectures. The study evaluated Vector RAG, Tree Reasoning, and a proposed Adaptive Hybrid Retrieval (AHR) system across 1,200 SEC filings and other financial, legal, and medical documents. Using a four-tier query complexity benchmark and GPT-4 as a judge, the research found that no single paradigm dominates. Tree Reasoning achieved the highest overall score (0.900) and perfect 100% cross-reference recall, but Vector RAG won on multi-document synthesis tasks.
The key finding is that performance depends heavily on query type and document structure. On the expert-annotated FinanceBench of 150 real SEC filing questions, the gap between methods widened: Tree Reasoning scored 0.938, Hybrid AHR scored 0.901, and Vector RAG scored 0.821—an 11.7 percentage point difference favoring tree-based approaches in real-world scenarios. This quantifies a critical capability gap in pure vector search for complex, structured document queries. The paper argues convincingly for the development of adaptive systems that can dynamically route a query to the best retrieval strategy (vector, tree, or hybrid) based on its assessed complexity.
All code and data from the study are publicly available, providing a valuable benchmark and toolkit for developers building enterprise-grade RAG systems for domains like finance and law where accuracy is paramount. The framework's tiered approach to query classification offers a practical blueprint for moving beyond one-size-fits-all retrieval.
- Tree Reasoning retrieval achieved a 0.938 score on FinanceBench, beating Vector RAG by 11.7 percentage points on real SEC filings.
- The proposed Adaptive Hybrid Retrieval (AHR) system excelled on cross-reference (0.850) and multi-section queries (0.929), showing the value of dynamic routing.
- Cross-reference recall hit 100% for tree-based and hybrid methods versus 91.7% for vector search, highlighting a key weakness in standard RAG.
Why It Matters
This research provides a blueprint for building more accurate, adaptive RAG systems for critical domains like finance and law, moving beyond one-size-fits-all vector search.