Enterprise RAG pipelines often call an LLM at every step, even for simple questions. This article from Towards Data Science shows how to save cost and latency by routing easy questions away from the model, using a cheap signal already in the pipeline.

The pipeline described in Article 9 (a production RAG system for PDFs) uses three model calls per question: one to parse, one to arbitrate retrieved candidates, and one to generate the answer. On hard questions, this ensures trustworthiness. But on easy questions like “What is the annual premium?” the model calls are unnecessary—a deterministic keyword match already isolated the answer.

The router uses the co_occurrence_score (from Article 7) computed on line_df to decide. If the top line’s score is high enough (≥4) and the margin over the runner-up is wide (≥3), the question is routed to a fast path that extracts the answer without any model call. This saves about two seconds per easy question. If the scores are flat (e.g., multiple lines tied at 2), the question goes through the full pipeline.

The threshold is a business decision, tuned per corpus. The key insight: the signal is already there—just read it and route. The companion notebook runs the router on a broker corpus, printing per-question confidence. This composes with the dispatcher from Article 6C.