EMPATHICED: AN EDUCATIONAL CHATBOT WITH RAG AND AFFECTIVE COMPUTING IN A COORDINATED SCIENTIFIC INITIATION CONTEXT

Abstract

Educational chatbots built on large language models face a persistent challenge: delivering fluent interaction without sacrificing factual accuracy, traceability, and institutional responsibility. At the same time, Undergraduate Research initiatives in Brazil are often organized around highly individualized models, which may constrain collaborative technological learning. This paper presents two complementary contributions. First, it introduces EmpathicEd, a modular architecture that combines a multi-stage Retrieval-Augmented Generation (RAG) pipeline with a socio-emotional layer based on multilingual BERT-like models, aiming to produce grounded, contextualized, and emotionally appropriate responses. Second, it proposes a replicable coordinated undergraduate research methodology organized around parallel teams, biweekly sprints, continuous documentation, and progressive pedagogical scaffolding over a ten-month period. The initial validation suggests promising results for both retrieval quality and emotional classification, while also making visible relevant methodological limitations such as the small pedagogical sample, the lack of full baselines, and the reliance on a proprietary institutional corpus. The paper argues that the combination of RAG, evidence governance, and socio-emotional adaptation is a relevant path for designing more trustworthy educational conversational agents in Brazilian higher education settings.