Document Segmentation in Legal Retrieval-Augmented Generation: An Empirical Study on Kazakh Legal Texts
Year: 2026
Type: conference-paper
Abstract: Retrieval-Augmented Generation (RAG) has become a practical approach for legal question answering by grounding generated responses in authoritative documents. In legal domains, retrieval performance is closely linked to how documents are segmented into retrieval units. This paper analyzes the role of document chunking in Legal RAG for Kazakh legal texts, a low-resource and morphologically rich language. We introduce a structure-aware semantic chunking pipeline that preserves legal hierarchy while applying semantic refinement within structural boundaries to control chunk length and coherence. The approach is evaluated under sparse, dense, and hybrid retrieval configurations on a curated Kazakh legal corpus. Results show that structure-aware segmentation yields more stable early ranking behavior, reflected in higher nDCG@10 and MRR, while maintaining competitive recall. Additional experiments indicate that cross-encoder reranking does not provide consistent gains in this setting and significantly increases latency. The findings emphasize egmentation as an important architectural component in Legal RAG systems for low-resource legal contexts.
Institutions: Astana IT University, Nazarbayev University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7168290051