Title: GRASP: Plan-Guided Graph Retrieval with Adaptive Fusion and Reranking on Semi-Structured Knowledge Bases

Abstract: Semi-structured knowledge bases (SKBs) organize textual content within typed graphs comprising entities and relationships, serving as the foundation for diverse applications including product search, academic literature discovery, and precision medicine queries. Current hybrid retrieval systems operating on SKBs typically face limitations, such as using the graph solely for query expansion, combining textual and structural data through global weighting schemes, or depending on fine-tuned graph-traversal generators. To address these challenges, we introduce GRASP, a novel three-stage retrieval framework for SKBs. This system integrates plan-based graph retrieval, plan-conditioned fusion utilizing a dense retriever, and a fine-tuned reranker applied to the combined candidates. GRASP achieves significant improvements in state-of-the-art performance, surpassing existing methods on all metrics across the three STaRK benchmarks. Notably, it increases the average Hit@1 score from 62.0 to 73.9. Additional ablation and sensitivity analyses validate the robustness and effectiveness of the GRASP approach.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC