Lightweight Query Routing for Retrieval-Augmented Dialogue: A TF–IDF and Logistic Regression Classifier with Confidence-Based Clarification

A lightweight five-way query router for RAG-based dialogue systems using TF–IDF + multinomial logistic regression. Routes user queries to NO_RAG, LOCAL_RAG, WEB_RAG, MEMORY_ONLY, or ASK_CLARIFICATION with a 0.45 confidence threshold fallback. Achieves 93.6% test accuracy, 0.935 macro F1, and 0.993 weighted ROC AUC on 470 balanced examples. Includes full pipeline, Streamlit demo, confusion matrices, and 5-fold CV.