Awesome Search KG

Lightweight Query Routing for Adaptive RAG: A Baseline Study on RAGRouter-Bench

3 min read

Lightweight Query Routing for Adaptive RAG: A Baseline Study on RAGRouter-Bench

Authors: Prakhar Bansal, Shivangi Agarwal Source: arXiv 2604.03455, April 2026

Summary

RAG strategies differ a lot in token cost: on RAGRouter-Bench, relative to an LLM-only answer, NaiveRAG costs 1.4×, GraphRAG 2.1×, HybridRAG 2.8× and IterativeRAG 3.5×. Most deployed systems still apply one strategy to every query. This short paper asks how far the cheapest possible routers get: classical classifiers over query text alone, predicting a query type that maps to a strategy.

RAGRouter-Bench (Wang et al., 2026) has 7,727 queries over four corpora — MuSiQue (Wikipedia, 3,356 queries), QuALITY (literature, 1,200), UltraDomain (legal, 1,277) and GraphRAG-Bench (medical, 1,896) — each labelled factual (52.9%), summarization (30.0%) or reasoning (17.1%). The benchmark’s own finding is that no single strategy dominates; the best one depends on the query and the corpus.

Setup

  • Five classical classifiers × three feature sets — TF-IDF, MiniLM sentence embeddings, and hand-crafted structural features — for 15 combinations.
  • Predicted type maps to a strategy following Adaptive-RAG: factual → NaiveRAG, reasoning → HybridRAG, summarization → IterativeRAG. The authors stress this mapping is a simplification for cost estimation only.

Findings

  • Best: TF-IDF + SVM, macro-F1 0.928, accuracy 93.2%.
  • Lexical beats semantic: TF-IDF (0.928) outperforms MiniLM embeddings (0.897) by 3.1 macro-F1 and structural features (0.788) by 14.0. Question-word type, domain terminology and patterns such as “summarize” or “compare” are enough to separate the three types; the authors’ likely explanation is that dense embeddings conflate surface-similar but type-different queries from different domains, while TF-IDF keeps the vocabulary signals that differ by type.
  • A simulated 28.1% token saving versus always running the most expensive strategy.
  • Medical queries are hardest to route (TF-IDF+SVM 0.803 macro-F1) — that corpus is a single long document, so every query type draws on the same text — and legal the most tractable (0.967).

Caveats

  • The savings are simulated from the type-to-strategy mapping, not measured answer quality under routing.
  • The classifiers see query text only; the benchmark shows corpus properties (connectivity, density, hubness and more) also decide which strategy wins, so these results are a floor, not a ceiling.
  • Predicting a query-type label is an easier task than predicting the best strategy directly.

Graph View

Backlinks