Enhancing RAPTOR with Semantic Chunking and Adaptive Graph Clustering

Yan liu¹Xiaodong Xie^1*Xin Wan¹Yi Pan²Cheng WANG¹

• ¹Huaqiao University School of Computer Science and Technology, Quanzhou, China
• ²School of Computer Science and Engineering, Changsha university, Changsha ,China, Changsha, China

While Retrieval-Augmented Generation (RAG) significantly enhances language models, its application to long documents is often hampered by simplistic retrieval strategies that fail to capture hierarchical context. The RAPTOR framework addresses this through a recursive tree-structured approach, yet its effectiveness is constrained by semantic fragmentation from fixed-token chunking and a static clustering methodology that is suboptimal for organizing the hierarchy. In this paper, we propose a comprehensive two-stage enhancement framework to address these issues. We first employ Semantic Segmentation to generate coherent foundational leaf nodes and then introduce an Adaptive Graph Clustering (AGC) strategy, leveraging the Leiden algorithm with a novel layer-aware dual-adaptive parameter strategy to dynamically tailor clustering granularity. Extensive experiments on both the narrative QuALITY benchmark and the scientific Qasper dataset demonstrate the robustness and domain generalization of our framework. Our full model achieves a peak accuracy of 65.5% on QuALITY and demonstrates superior semantic validity on Qasper, significantly outperforming the baseline. Furthermore, comparative ablation studies reveal that our graph-topological approach outperforms traditional distance-based, density-based, and distribution-based clustering methods. In addition to performance gains, our approach constructs a dramatically more compact hierarchy, reducing the number of required summary nodes by up to 76%. This work underscores the critical importance of a holistic, semantic-first approach to building more effective and efficient retrieval trees for complex RAG tasks across diverse domains. To facilitate future research and reproducibility, we have made our source code and data publicly available at: https://github.com/Xin5643/Graph-raptor.