Comparative Analysis of Static, Adaptive, and Agentic RAG Strategies for LLM-Based Online Tutoring
Year: 2026
Type: conference-paper
Abstract: This research-to-practice full paper describes a comparative analysis of Static, Adaptive, and Agentic Retrieval-Augmented Generation (RAG) strategies for large language model (LLM)-based online tutoring systems. Large language models (LLMs) are increasingly used in online tutoring systems, but ensuring factual accuracy and reliable reasoning remains a significant challenge. Retrieval-Augmented Generation (RAG) integrates LLMs with external knowledge sources, providing factual grounding for generated responses while avoiding the need for model parameter tuning. Various RAG strategies have been explored in the literature, including static single-pass retrieval, agentic multi-step retrieval, and query-classification-based adaptive retrieval. However, no prior work has comprehensively analysed and compared these strategies in terms of reasoning performance, retrieval quality, and computational cost within an educational question-answering context. Therefore, this work implements three RAG strategies: Static, Agentic, and Adaptive, and evaluates them based on (1) retrieval effectiveness metrics, (2) the RAG Triad framework for assessing answer quality, and (3) computational metrics within an LLM-based online tutoring system using the FairytaleQA and RACE educational question-answering datasets. Experimental results show that Adaptive RAG achieves the best balance between retrieval performance and computational efficiency, improving recall by 23.43% and mean reciprocal rank (MRR) by 21.8% compared with Static RAG, while introducing only moderate increases in latency and token consumption. Moreover, for the RACE dataset, Adaptive RAG reported relative improvement of 27.16% and 29.6% in recall compared to Static and Agentic RAG strategies, respectively. These findings highlight the importance of retrieval strategy design in developing reliable and scalable AI-driven tutoring systems and provide practical guidance for deploying cost-aware RAG-based educational applications.
Institution: University of Glasgow
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7171269786