SLLM을 활용한 RAG기술 기반의 교육기관 문서 질의응답 기법 활용 방안

초록

Recent advances in large language models (LLMs) have significantly accelerated the adoption of natural language technologies in educational settings. Retrieval-Augmented Generation (RAG) further enhances the capabilities of LLMs by dynamically retrieving up-to-date or domain-specific knowledge from external sources, enabling more accurate and trustworthy responses. Nevertheless, the direct deployment of LLMs in educational institutions is often limited by high inference costs and privacy concerns. Moreover, institutional documents such as admission guidelines or academic regulations typically contain domain-specific terminology, deeply hierarchical structures, and highly formatted content including merged-cell tables and embedded images. Conventional LLM-based chatbots struggle with accurately extracting data from complex tabular layouts, effectively identifying relevant paragraphs within long, multi-level documents, and incurring substantial computational costs for each query. To address these limitations, this study proposes a framework that combines a Small Language Model (SLLM) with RAG. We introduce a section-based document-parsing strategy that systematically decomposes institutional documents and a table-aware encoding pipeline that allows the SLLM to handle intricate tabular data accurately. The proposed approach offers a cost-effective and privacy-conscious alternative to commercial LLM services while maintaining high response accuracy for document question-answering tasks in educational domains.

키워드