Competence-Aware Retrieval-Augmented Generation for AI-Driven Tutoring in Software Engineering Education

Artificial Intelligence (AI) driven tutoring is increasingly used in digital and hybrid Software Engineering (SE) education to provide real-time feedback and explanations to students without the need for human tutors. However, many current approaches rely on generic Large Language Models (LLMs) that are not sufficiently tailored to the course materials and do not adapt to students' individual competence levels and preferences. As a result, the responses may lack contextual relevance and pedagogical differentiation. In this paper, we present a competence-aware Retrieval-Augmented Generation (RAG) approach for AI-driven tutoring in SE education, grounded in a requirements analysis involving 34 students. Based on the identified requirements regarding trust, task support, and adaptive feedback, we developed a system that automatically processes course materials, such as analyzing lecture slides, documents, and videos, and enriches them through summarization, semantic tagging, and cross-material linking. The processed resources serve as a knowledge base for an AI-tutor capable of answering course-specific questions and referencing course material via semantic search. The AI-tutor adapts the depth and structure of explanations to the students' competence level, motivational preferences, and assessment conditions. We implemented the AI-tutor in MEITREX, an intelligent tutoring system for SE education, and evaluated it in a user study with 17 students using a dedicated SE course environment. The results show that reference-based responses contributed positively to perceived trust and usability, while competence-aware and proactive mechanisms revealed a need for further refinement. The findings highlight design aspects for integrating competence-aware RAG into SE education.以人工智能（Artificial Intelligence，AI）为驱动的辅导系统正日益广泛应用于数字化与混合式软件工程（Software Engineering，SE）教育场景，可为学生提供实时反馈与讲解，无需依赖人类导师。然而，当前多数相关方法依赖通用型大语言模型（Large Language Model，LLM），此类模型未针对课程内容进行充分定制，也无法适配学生的个体能力水平与学习偏好，由此生成的回复可能缺乏语境关联性，且未体现教学法层面的差异化适配。本文针对SE教育中的AI驱动型辅导场景，提出一种能力感知型检索增强生成（Retrieval-Augmented Generation，RAG）方法，该方法基于针对34名学生开展的需求分析构建。基于明确的信任、任务支持与自适应反馈三类需求，我们研发了一套可自动处理课程资源的系统：该系统可对课件幻灯片、文档及视频等课程材料进行分析，并通过摘要生成、语义标注与跨资源关联等方式对其进行增强处理。经处理后的资源将作为知识库，供AI辅导系统使用，该系统可通过语义搜索匹配课程材料，回答与课程相关的专属问题。该AI辅导系统可根据学生的能力水平、学习动机偏好与考核场景，调整讲解的深度与结构。我们将该AI辅导系统集成至面向SE教育的智能辅导系统MEITREX中，并依托专属的SE课程环境，联合17名学生开展了用户评估研究。评估结果显示，基于参考资料生成的回复对提升用户感知的信任度与系统可用性具有积极作用，但具备能力感知与主动响应机制的模块仍有待进一步优化。本研究结果为在SE教育场景中集成能力感知型RAG方法提供了设计参考方向。