← Back to concept explainers

Concept Fable

Multi-hop retrieval

Conclusion first:Multi-Hop Retrieval is multi-step retrieval performed to answer complex questions: first find one piece of evidence, then continue retrieving based on the entities, times, relationships, or intermediate conclusions within it. It applies to tasks that require an evidence chain, such as contracts, knowledge graphs, incident analysis, and enterprise policy Q&A. The key risks are wrong hops, evidence drift, and rising cost, so hop control, query rewriting, chain logging, and final citations are needed.

Multi-Hop Retrieval · Retrieval

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

The truth about the debt that one index card could not find

The caravan accountant received a question: why did the North Gate customer refuse to pay for the third batch of goods? The new apprentice first checked the index and found only one refusal letter, which said "goods arrived late." Customer service was about to explain it as a delay, but the old accountant frowned. The refusal letter only stated the outcome—it didn't say why the goods were late, nor whether the customer had the right to deduct payment. The new apprentice lengthened the keywords and kept searching for "North Gate, third batch, late arrival." A few similar letters appeared on the desk, but they all stopped at the same surface-level facts.

The old accountant followed the trail down the chain: the refusal letter pointed to the delivery note, the delivery note pointed to the transport record, the transport record pointed to a temporary reroute, and the reroute pointed to a warehouse safety inspection.

In the end the answer changed. The client withheld payment not only because of the late arrival, but because a safety inspection forced a reroute, and after the reroute the caravan did not give advance notice as required by the contract, so the client had the right to deduct part of the payment.

The accountant wrote this process into the rules. Every hop must explain why the next piece of evidence is needed, and what person, time, or relationship the previous piece of evidence provided. When evidence is insufficient, it is better to stop than to force a complete chain.

Doing it this way is slower than flipping through the book once, and more likely to drift off course midway, so the number of skips must be limited, the evidence must leave a trace, and the conclusion must be traceable back to every original document.

Only then did the new apprentice learn to build a chain of clues. For every page found, he had to ask which passage it answered and which page it pointed to next; if a link in the middle was broken, the final explanation could only be regarded as a guess and could not be used to bill a client or assign responsibility.

The old accountant also required the new clerk to write every jump on the margin. Why the refusal letter pointed to the delivery note, and why the delivery note pointed to the shipping record, all had to make sense. If a jump existed only because the words looked alike, it could not enter the final answer. Only after the trail was complete did the answer change from "late arrival" to "a safety inspection triggered a deductible clause." Later the accountant set a rule for the new clerk: any answer that could only hold by crossing a page must bring every jump back to the table. Speaking from only the last page looks fast, but it often loses the intermediate causes.

Later, when the caravan handled complex questions, it no longer expected a single sheet of paper to solve everything. The real answer is often hidden in the paths between pieces of evidence.

Reveal

This story is about: multi-hop retrieval

Multi-Hop Retrieval is multi-step retrieval performed to answer complex questions: first find one piece of evidence, then continue retrieving based on the entities, times, relationships, or intermediate conclusions within it. It applies to tasks that require an evidence chain, such as contracts, knowledge graphs, incident analysis, and enterprise policy Q&A. The key risks are wrong hops, evidence drift, and rising cost, so hop control, query rewriting, chain logging, and final citations are needed.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Chargeback email: first-hop evidence
  • Delivery notes, transport records, safety inspections: subsequent retrieval jump points
  • Each hop explains why: the reasoning and retrieval chain of multi-hop retrieval
  • Limiting hop count: cost and drift control
  • Paths between evidence: the true answer structure of complex problems

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

Graph RAGreasoning
京ICP备2026014272号