Add a router that skips retrieval when your documents won't help

Why does typing 'hi' to your RAG chatbot still trigger a full document search?

IntermediateRead 5 minDoing: an hour or two, once your RAG pipeline already answers questions end to end7 steps

What you'll have at the end

A routing step that answers straight from the model for questions it doesn't need documents for, tested on five real ones

Before you start

You need

A RAG pipeline that already retrieves and answers from your documents for every question, with access to call the same LLM the rest of the pipeline uses for one quick classification prompt.

Not covered

Splitting retrieval further into single-hop versus multi-hop pipelines; this only covers the one decision of retrieving at all or skipping straight to an answer.

Leans on

7 steps

Where to go next

Make your RAG system say it does not know instead of guessing

when the no-retrieval branch itself starts guessing at something it doesn't actually know, rather than the routing decision being wrong

Your RAG's retrieval hit rate has never been measured

to score the router's own accuracy for real once you have more than five questions to check it against

Checked 18 Aug 2026

Part of the RAG (Retrieval Augmented Generation) cookbook