In recent years, large language models (LLMs) have demonstrated outstanding performance in natural language processing tasks, but still face issues such as knowledge lag, hallucination generation, and limited understanding of domain knowledge, which restrict their application in highly reliable educational scenarios. The retrieval-enhanced generation (RAG) technology alleviates these problems to some extent by introducing external knowledge sources. However, most existing RAG systems in the education field still focus on improving the accuracy of answer generation, lacking active guidance of the students’ cognitive process and the integration of systematic teaching strategies. To address this deficiency, this paper proposes a retrieval-enhanced generation intelligent teaching system with the ability to actively guide thinking. The system firstly identifies the user's question type using a lightweight BERT classifier and selects corresponding retrieval and teaching guidance strategies accordingly; secondly, it improves the quality of the context through semantic retrieval and cross-encoder reordering mechanisms; thirdly, it designs four types of teaching guidance templates to achieve differentiated responses; finally, it generates responses with both accuracy and educational significance based on the DeepSeek model. Experimental results show that this system demonstrates good performance in terms of reasoning efficiency, retrieval quality, and teaching guidance ability.