Accessible and Reliable AI Coding Tutors: Augmenting Large Language Models with Retrieval-Augmented Generation for Java Programming

This paper addresses the challenge of improving the reliability and accuracy of Large Language Models (LLMs) for assisting students in learning Java programming, a critical component of object-oriented computer science courses. While LLMs have shown promise in generating code, they often produce incorrect or unreliable outputs, which can hinder the learning process. To mitigate these issues, we propose a novel augmentation framework that integrates Retrieval-Augmented Generation (RAG), enabling LLMs to retrieve best-practice Java code examples from external sources to enhance the accuracy of generated solutions. We evaluate this framework using 250 Java coding tasks, covering a range of difficulties. Our findings show that the RAG-augmented model, when applied to a lightweight LLM (Google DeepMind's Gemma), outperformed baseline model by 19 % for mid and high-difficulty problems. In particular, Augmented Gemma generated accepted code for 11 problems where no other model could provide a valid solution. This suggests that retrieving external best-practice examples is critical in addressing complex coding challenges. These results highlight the potential of lightweight, accessible models enhanced with RAG to provide reliable AI coding assistance in educational settings, facilitating both accurate problem-solving and reinforcement of best coding practices.