Split a multi-hop question into sub-questions RAG can answer

One RAG answer needs facts scattered across two different documents, and a single search only turns up one of them.

Intermediate to advancedRead 6 minDoing: a few hours8 steps

▶ Start

What you'll have at the end

A working example where the two-hop question gets answered correctly after failing under plain single-shot retrieval

Before you start

You need

A working RAG pipeline (a retriever plus a generation step) that already answers single-fact questions correctly, and an LLM you can call a second time mid-request to judge and generate follow-up questions.

Not covered

A dedicated query-complexity classifier that routes questions across several pipelines before they arrive here. This recipe's short-circuit lives inside the self-ask prompt itself.

Leans on

8 steps

Where to go next

Your RAG's retrieval hit rate has never been measured

measure whether decomposition actually raised retrieval hit rate across your real traffic, since this recipe only proves it on one worked case.

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

if most incoming questions never need retrieval at all, route those away before they ever reach the self-ask check this recipe adds

Checked 18 Aug 2026

Part of the RAG (Retrieval Augmented Generation) cookbook