multi-hop retrieval

Multi-hop retrieval follows chains of reasoning across multiple document retrievals to answer complex questions. **Problem**: Some questions require information from multiple documents that must be connected logically. "Who founded the company that made the device used in the Apollo missions?" **Mechanism**: First retrieval answers partial question → extract entities/facts → formulate follow-up query → retrieve again → chain until complete. **Approaches**: **Iterative**: Retrieve → reason → retrieve again based on findings. **Query decomposition**: Break complex query into sub-queries, retrieve for each, synthesize. **Agentic**: Agent decides when more retrieval needed and what to retrieve. **Example flow**: Q: "CEO of company that acquired Twitter" → retrieve "Elon Musk acquired Twitter" → retrieve "Elon Musk is CEO of Tesla, SpaceX" → answer. **Challenges**: Error accumulation across hops, determining when to stop, increased latency. **Evaluation**: Multi-hop QA benchmarks (HotpotQA, MuSiQue). **Frameworks**: LangChain multi-hop retrievers, custom agent loops. **Optimization**: Cache intermediate results, limit hop depth, verify reasoning chain. Essential for complex reasoning over knowledge bases.