Summary
Multi-hop retrieval: some questions require chaining retrieved facts together. Iterative retrieval — retrieve, read, generate a sub-question, retrieve again — solves these; single retrieval passes cannot. Frameworks: DSPy, LlamaIndex query engines, agentic RAG.
Objectives
- 01Recognize multi-hop questions.
- 02Describe the iterative retrieval loop.
- 03State the accuracy/latency tradeoff.
The Lesson
The pattern
'Who directed the movie that won Best Picture in 1974?' requires two retrievals: identify the movie, then identify the director. A single-pass retrieval finds either fact but rarely both. Loop: retrieve → answer sub-question → generate next query → retrieve → repeat.
The cost
Each hop adds latency (retrieval + generation of the next query). Cap at 3-5 hops to prevent runaway loops. Instrument every hop; the failure mode is 'model asks the same question three times.'
Key Ideas
- Some questions cannot be answered in one retrieval; loop is the answer.
- Cap hops and instrument every one.