Agent educatiu multiassignatura: disseny i implementació amb Rasa i Docker

Abstract

[en] Nowadays, general interest on chatbots or Conversational Agents has augmented drastically due to the introduction of the Large Language Models (LLMs) empowering ChatGPT and Google Bard. Despite the high intellectual capacities and use of natural language, LLMs suffer defects such information deliriums and loss of the context window of the conversation, making them unreliable for specific tasks in some controled environments. This project presents a Conversational Agent provided by Rasa and enhanced with calls to OpenAI’s GPT-3.5-turbo model to perform specific NLG (Natural Language Generation task) complementing the Rasa intents-driven design. Project context is a multi-subject (Programació I i Mètodes Numèrics) learning support Conversational Agent for the degrees of Computer Engineering and Mathematics at the Universitat de Barcelona, designed to assist the student with doubts about the teaching plan of the subjects and with general conceptual doubts. It also implements the delivery of exercises for students to solve, along with secure execution systems for C and Java languages through Docker. To facilitate the implementation of the mentioned exercises and their tests, a unit test library for the C language for teaching staff is presented, along with various examples. Some of the chosen examples are well-known numerical methods which have been expanded, extended, and theoretically formalized. To enable the agent to perform the described functionalities, an architecture prepared for deployment via Docker-Compose is presented. This allows the agent to store and access specific data, update these in real time, execute corrections of the exercises, communicate with users through the Telegram interface, and keep records of the conversations per student for subsequent analysis.