Multi-hop RAG: follow evidence across connected documents

When one passage cannot answer the question

Multi-hop RAG gathers evidence through more than one retrieval or relationship step before a model answers. The intermediate step connects the question to evidence that may use different names.

Consider a project move. Its project note identifies an access permit. The permit points to a checklist and an approver. Answering who approved the move's permit and what must be ready requires both branches.

The GraphRAG foundation explains the broader family of graph retrieval methods. This article examines the evidence needed to complete one branched question.

Follow Project Harbor through four original sources

We created four fictional documents for this example:

Project Harbor
  depends_on → Cedar access permit
    depends_on → Safety checklist → clear loading bay
    decided_by → Imani Cole → dated approval source

This diagram is a reading aid. The downloadable fixture supplies the actual authored graph proposal and exact source passages.

The question is: "Who approved the permit for Project Harbor, and what must be ready before arrival?" An answer needs the approval and checklist evidence.

Inspect the graph step we executed

Against the Full 1.1.0 archive, direct graph expansion reached both terminal sources from the Project Harbor entity. Each terminal result carried a two-edge path and a quotation that passed source verification.

The test supplied no initial retrieval candidates. It isolates graph traversal from keyword ranking, vector ranking and generation. It does not benchmark a complete RAG answer.

Open the sources, paths, receipts and configuration. The record identifies the archive, runner, fixture, runtime and execution date.

Break the bridge while keeping the final sources

The fixture then changed the disposable permit document so it no longer supported the approval or checklist connections. Both terminal sources disappeared from that traversal.

The checklist and approver files still existed. Their facts did not become false merely because the bridge changed. The previous path no longer supplied current support for reaching them from this project.

This failure differs from a missing source file or an incorrect quotation. Review each intermediate relation as well as the final passage.

The fixture uses deliberately authored relationships. It does not test automatic extraction or a model deciding which relationship is true.

Distinguish expansion from reranking and context assembly

Reranking reorders existing candidates. Expansion can add evidence through graph connections. Context assembly packages a bounded selection for the agent's task.

In a complete workflow, a model still needs to combine the two facts without inventing a third. A clear loading bay requirement does not establish that the bay is already clear.

Graph paths can also grow too broad. Record the allowed relation types, project boundary, traversal budget and missing evidence. A path that stops at a limit may be incomplete.

Use the four-mode comparison for executed retrieval rankings. Keep that experiment separate from this direct graph-stage demonstration.

Repeat the failure before relying on the answer

1. List every source required for the answer.
2. Inspect the intermediate relationships.
3. Verify the returned quotations.
4. Change one bridge in a disposable fixture.
5. Repeat the traversal.
6. Record the missing path and any remaining evidence.
7. Test a generated answer separately.

Download the multi-hop review and reproduction instructions. Continue with citation verification and evaluation.

Check your integration before choosing a Gigabrain package.

Method note: Gigabrain Labs prepared this article with AI assistance from its original fictional fixture and executed component output.