Authors: Korade Apeksha Maruti, Professor Bangar Abhishek Popat, Professor Bhosale S.B.
Abstract: Learners often spend a great deal of time condensing lecture slides, textbook chapters and research papers into brief revision notes. Large language models (LLMs) from the generative AI family can draft such notes within seconds, but a model that is not grounded in a source may write confident text that the material never states. Retrieval-augmented generation (RAG) lowers this risk by supplying the model with passages fetched from the learner’s own documents. This paper offers a focused narrative review of verified literature on AI-driven study notes generation with generative AI and RAG. Three linked areas are covered: retrieval and RAG techniques (chunking, dense embeddings, retrievers and advanced pipelines); generative models and prompting (LLMs, structured prompt frameworks and hallucination); and empirical evidence on generative AI in engineering and computing courses. The reviewed work indicates that RAG was proposed to anchor generation in retrieved text, that dense retrieval outperformed a strong sparse baseline in open-domain question answering, and that in a quasi-experiment with 26 students, structured prompts produced fuller answers than free-style prompts. Even so, most educational studies capture student perceptions instead of learning outcomes, rely on small samples, and none of the reviewed studies directly measures note quality, faithfulness or retention. The paper contributes a comparison of representative studies, a conceptual seven-layer architecture that links ingestion, retrieval, grounded generation, verification and learner feedback, and a ranked list of research gaps.
