Classical Query Routing for Adaptive Retrieval-Augmented Generation: Reproducibility, Robustness, and Cost-Aware Evaluation

This record contains the preprint and LaTeX source for an auditable reproduction and non-deep-learning extension of query routing on the public RAGRouter-Bench dataset. The study evaluates protocol sensitivity, source-order and grouped cross-validation, leave-one-domain-out robustness, leakage controls, classical feature representations, and an illustrative routing-cost simulation. The simulated cost is not API billing, latency, retrieval quality, or generation quality. The system predicts a route but does not execute the complete retrieval-and-generation pipeline. Code and reproducibility materials:https://github.