Conflict-Aware RAG Routing: Balancing Cost and Accuracy via Context Contradiction

Abstract

Adaptive model routing can reduce the cost of retrieval-augmented generation (RAG), but the routing signal must identify when a small language model (SLM) is likely to fail. We study context conflict, the contradiction among retrieved passages, as one such signal. A five-feature Random Forest combines conflict with query length, retrieved-context length, chunk overlap, and Reciprocal Rank Fusion (RRF) confidence. We evaluate the pipeline on 3,000 multi-hop questions from MuSiQue, HotpotQA, and 2WikiMultiHopQA using Qwen2.5-1.5B-Instruct locally and DeepSeek-V4-Flash through an API. Nested five-fold calibration selects budget cutoffs from inner out-of-fold rather than in-sample training scores. At an approximately 65% LLM-call rate, the hybrid router saves about 35% of API calls while retaining 83.2–92.2% of the Always-LLM F1. Its advantage over RRF alone is dataset-dependent: at the 50% budget it changes F1 by −1.08, −0.03, and +1.48 points on MuSiQue, HotpotQA, and 2Wiki, respectively; only the 2Wiki interval excludes zero. At 30%, it offers no consistent advantage. An exploratory high-RRF/high-conflict subgroup analysis finds larger hybrid-versus-RRF gains on HotpotQA and 2Wiki. However, a direct additive conflict override at +2/5/10 usage points does not outperform equal-cost random escalation. Thus, contradiction can be complementary on some retrieval distributions, but a simple conflict override is not generally justified.