Structural Chunking: A Semantic-Structural Integrated Method for Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) systems rely heavily on the quality of document chunking, which determines the granularity and contextual continuity of retrievable units. Existing chunking methods inevitably trade off between semantic coherence and computational efficiency. To overcome this limitation, we propose Structural Chunking, a novel paradigm that integrates semantic cohesion with structural consistency by leveraging both surface- and physical-level document features. Unlike purely semantic methods, Structural Chunking quantifies structural patterns and fuses them with semantic embeddings to compute a semantic-structural cohesion score. Chunk boundaries are then detected where this composite distance sharply increases. Experiments on five BEIR benchmark datasets demonstrate that Structural Chunking achieves the most consistent recall and average precision across domains, outperforming semantic and statistical baselines while maintaining stable chunk-size distributions. The results indicate that incorporating structural hierarchy into the RAG preprocessing pipeline substantially enhances retrieval efficiency and contextual fidelity, suggesting a new direction for structure-aware information retrieval.