Reproduce “Lightweight Query Routing for Adaptive RAG: A Baseline Study on RAGRouter-Bench” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Prakhar Bansal, Shivangi Agarwal

Researchers from IIIT Delhi and an independent researcher introduce a lightweight, classifier-based system for adaptive query routing in Retrieval-Augmented Generation (RAG) pipelines, achieving up to 28.1% token savings with high accuracy using lexical features for classifying query types. The study establishes initial baselines for query-side routing on the RAGRouter-Bench dataset, highlighting the varying effectiveness of different feature sets and classifiers.

Key insights

  • Lexical features (TF-IDF) were found to be more effective for query-type classification than semantic (MiniLM) or hand-crafted structural features on the RAGRouter-Bench dataset.
  • Optimal RAG paradigm selection is influenced by both query type and corpus-side signals, suggesting that query-text-only routing provides a strong baseline but has inherent limitations.
  • There is a critical trade-off between maximizing token savings and maintaining routing accuracy; optimizing for cost alone can severely degrade the quality of generated responses.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.