Skip to main content
Xuhang Shi
All projects

EvLink: Source-grounded Evidence Linking for Graph RAG

A Graph RAG retriever that only hops between passages along links a source passage actually supports, then picks a compact evidence set that covers every part of a multi-hop question.

EvLink overview on a two-hop question: evidence-link construction, coarse-grained evidence induction over a query-local subgraph, and fine-grained evidence retrieval with evidence-need coverage.

Highlights

  • Outperforms leading GraphRAG baselines on five QA benchmarks, with average gains of +2.4 R@5, +1.9 EM and +2.4 F1.
  • Connects passages only through transitions grounded in a source passage, not through semantic similarity alone.
  • Runs end to end, or as a drop-in evidence selector on top of an existing dense, sparse or graph retriever.

Problem

Graph RAG organises a corpus into a graph so that a retriever can hop from passage to passage. But an edge in that graph usually means two passages are related, not that one justifies the next step of the reasoning. A passage can be one hop away and still fail to support the transition the question needs, so the reader ends up with distractors instead of the bridge passage.

Approach

EvLink keeps passages as the units of retrieval and only connects them through transitions that a source passage supports:

  • Evidence links. Relation-grounded links come from an explicit relation stated in a source passage; endpoint-aligned links act as a source-bounded fallback.
  • Coarse-grained evidence induction. Starting from the question’s anchors and dense seed passages, a bounded breadth-first search over evidence links builds a query-local subgraph and recovers bridge passages that similarity search misses.
  • Fine-grained evidence retrieval. The question is broken into evidence needs (in the example: which river does Darling Mills Creek lead to? and what does that river flow into?), and noisy-OR coverage selects a small, non-redundant set of passages that satisfies every need within a fixed reader budget.

Results

On three multi-hop benchmarks (HotpotQA, 2WikiMultiHopQA, MuSiQue) and two single-hop benchmarks (Natural Questions, PopQA), EvLink consistently outperforms leading GraphRAG baselines, with average gains of 2.4 R@5, 1.9 EM and 2.4 F1.

Software

EvLink is released as an MIT-licensed Python package (evidencelink) with explicit artifact schemas, an offline and deterministic end-to-end example, and inspectable selection traces. It can run the full pipeline or select evidence from the ranked candidates of a retriever you already use.