Multi-Hop Reasoning

Answering questions that require traversing multiple relationship steps through a knowledge base, where no single document contains the complete answer.

“What regulations affect this product in these markets?” requires connecting: product → applicable standards → regulatory bodies → geographic jurisdictions. Each hop crosses an entity boundary.

Vector similarity can’t do this, it finds content that resembles the question, not content connected by unstated relationships. You’d need the answer to already mention all the intermediate concepts in a single chunk.

Multi-hop reasoning requires explicit relationship structure: knowing that A relates to B, B relates to C, therefore A can reach C through a path. This is the core argument for graph-based retrieval over pure vector search.

The challenge is that real questions mix single-hop (direct lookup) and multi-hop (traversal) needs. Systems optimized for one often underperform on the other.

Related: 07-atom—directness-comprehensiveness-tradeoff, 07-molecule—vectors-vs-graphs