Reframing classic versus agentic as a per-query routing decision rather than a global architecture choice is the right move, and a classifier at 93.2% accuracy cutting 28.1% of tokens is a convincing enough number to act on. The part I would stress-test is the asymmetry of the routing errors, since sending a genuinely multi-hop query down the cheap single-pass path costs you a correct answer, while the reverse only costs tokens. Did the RAGRouter-Bench numbers report routing accuracy per query type, because the tail of hard queries is where that 6.8% miss rate actually bites?
Reframing classic versus agentic as a per-query routing decision rather than a global architecture choice is the right move, and a classifier at 93.2% accuracy cutting 28.1% of tokens is a convincing enough number to act on. The part I would stress-test is the asymmetry of the routing errors, since sending a genuinely multi-hop query down the cheap single-pass path costs you a correct answer, while the reverse only costs tokens. Did the RAGRouter-Bench numbers report routing accuracy per query type, because the tail of hard queries is where that 6.8% miss rate actually bites?