Relink: Constructing Query-Driven Evidence Graph On-the-Fly for GraphRAG

TL;DR

Relink challenges GraphRAG's reliance on static knowledge graphs by introducing a reason-and-construct paradigm that builds a query-specific evidence graph on the fly. It fuses a high-precision explicit KG backbone with a high-recall latent relation pool derived from corpus co-occurrences and learns a unified semantic space with a query-driven ranker to select edges that directly support the query. Through iterative path expansion, dynamic latent relation instantiation, and evidence-grounded answer generation, Relink repairs incomplete reasoning chains and filters distractors, producing faithful answers. Experiments on five open-domain QA benchmarks show consistent improvements in EM and F1, along with strong robustness to knowledge sparsity and improved reasoning traceability.

Abstract

Graph-based Retrieval-Augmented Generation (GraphRAG) mitigates hallucinations in Large Language Models (LLMs) by grounding them in structured knowledge. However, current GraphRAG methods are constrained by a prevailing \textit{build-then-reason} paradigm, which relies on a static, pre-constructed Knowledge Graph (KG). This paradigm faces two critical challenges. First, the KG's inherent incompleteness often breaks reasoning paths. Second, the graph's low signal-to-noise ratio introduces distractor facts, presenting query-relevant but misleading knowledge that disrupts the reasoning process. To address these challenges, we argue for a \textit{reason-and-construct} paradigm and propose Relink, a framework that dynamically builds a query-specific evidence graph. To tackle incompleteness, \textbf{Relink} instantiates required facts from a latent relation pool derived from the original text corpus, repairing broken paths on the fly. To handle misleading or distractor facts, Relink employs a unified, query-aware evaluation strategy that jointly considers candidates from both the KG and latent relations, selecting those most useful for answering the query rather than relying on their pre-existence. This empowers Relink to actively discard distractor facts and construct the most faithful and precise evidence path for each query. Extensive experiments on five Open-Domain Question Answering benchmarks show that Relink achieves significant average improvements of 5.4\% in EM and 5.2\% in F1 over leading GraphRAG baselines, demonstrating the superiority of our proposed framework.

Figures (4)

• Figure 1: Static GraphRAG failures vs. Relink's Dynamic Construction. Pre-built knowledge graphs cause two critical failures in GraphRAG: (a) missing links breaking reasoning paths, and (b) distractor facts (query-relevant but goal-misaligned). In contrast, our reason-and-construct approach, Relink, addresses both by discarding distractor facts and dynamically instantiating missing ones from the latent relations derived from the original text corpus.
• Figure 2: Relink's dynamic evidence graph construction. Relink iteratively builds reasoning paths by leveraging candidates from both the explicit KG ($\mathcal{G}_b$), and the latent co-occurrence relation pool ($\mathcal{R}_c$) derived from the corpus. Encoders $E_L$ and $E_F$ project these candidates into a unified semantic space where a query-driven ranker evaluates their relevance. When latent relations are selected, an LLM instantiates them into factual relations (e.g., "manuscript resides in") using source context, dynamically repairing missing path segments during construction.
• Figure 3: Performance trend as the factual graph is reduced. Relink exhibits remarkable robustness to knowledge sparsity, whereas the baseline's performance collapses.
• Figure 4: A case study contrasting static reasoning with Relink's dynamic approach. The static baseline (w/o$\mathcal{R}_c$) is misled by the highly relevant resides in distractor. In contrast, Relink succeeds by dynamically constructing the correct reasoning chain (composer of$\rightarrow$born in) and using its query-driven ranker to prioritize it.