Retrieval over semi-structured knowledge bases (SKBs) is tricky because these systems mix free-text descriptions with explicit relational structure. A good retriever must respect both, but most existing methods lean heavily on text. That works for capturing the general meaning of a query, yet it falls short when the query requires following chains of relations across multiple steps.
A new arXiv paper, GRASP (arXiv:2605.30237), proposes a plan-guided approach to graph retrieval. The method is designed to handle the relational constraints that text-based retrievers miss, using adaptive fusion to combine different sources of evidence and reranking to improve the final results.
The paper's abstract highlights the core problem: text-based retrievers capture semantics but not multi-hop relations. GRASP appears to address that gap directly, though the full details of the method and its evaluation are not included in the abstract. As the only source, this summary reflects the paper's stated motivation and proposed solution. Further reading of the full text would be needed to assess how well GRASP performs in practice.