Apontamento e Configuração de um Tutor com Inteligência Artificial (TIA) : Uma Análise Comparativa entre o Uso Direto do LLM e RAG no Ensino de Cálculo I

This work presents the development of an AI-powered tutor employing a Socratic approach to support the teaching of Calculus I, comparatively evaluating two implementation architectures: direct interaction with a Large Language Model (LLM) and integration of the LLM into a Retrieval-Augmented Generation (RAG) pipeline combined with prompt engineering. The objective was to analyze how these different approaches impact the tutor’s ability to adopt a contextualized Socratic pedagogical stance, guiding students in problem-solving without providing direct answers. The methodology encompassed the implementation of both solutions using the GPT-4 model (OpenAI), followed by simulations with a set of 27 Calculus I questions to evaluate each approach’s performance. The tutor’s interactions were recorded and analyzed for adherence to Socratic practices. The results showed that the RAG architecture with the few-shot learning technique achieved approximately 93% compliance with the expected Socratic style, whereas the RAG pipeline with one-shot achieved about 11%. Direct use of the LLM served as a comparative baseline, demonstrating that RAG combined with prompt engineering techniques provides an effective mechanism to regulate the LLM’s behavior, enabling the intelligent tutor to reliably replicate Socratic teaching strategies. The study demonstrates the feasibility and advantages of this approach for the development of virtual tutors in education, although future empirical evaluations with users are necessary to confirm the observed pedagogical gains.