What is Multi-Hop Retrieval?
Multi-hop retrieval is a technique where answering a query requires multiple sequential retrieval steps, with each step using information from previous retrievals to guide the next search. Rather than retrieving all relevant information in a single query, multi-hop retrieval breaks complex questions into sub-questions, retrieves information for each, and chains the results together to form a complete answer. This approach is essential for questions that require synthesizing information from multiple sources or following chains of reasoning.
The process typically involves decomposing a complex query into simpler sub-queries, retrieving information for the first sub-query, using those results to formulate the next sub-query, and continuing this process until sufficient information is gathered to answer the original question. For example, answering "Who was the president when the company Alice works for was founded?" might require first retrieving information about Alice's employer, then retrieving information about when that company was founded, then retrieving who was president at that time.
Multi-hop retrieval is particularly important for complex analytical questions, research tasks, and scenarios involving knowledge graphs where information is distributed across multiple documents or nodes. It's often implemented in agentic RAG systems where the agent can dynamically decide when and how to perform additional retrieval steps. GraphRAG naturally supports multi-hop retrieval through graph traversal, while document-based systems implement it through iterative query refinement and chaining. The approach increases system complexity but enables handling of questions that single-step retrieval cannot adequately address.