Database Tutor Agent: Integrating LLM and Knowledge Graphs for Effective Guidance in Relational Algebra and SQL Practices

Abstract

This paper presents the development of a tutor agent based on large language models (LLM) and knowledge graphs (KG), implemented in graph-oriented databases. This agentic flow is dedicated to assisting engineering students in the field of databases. It may be expanded to include other subjects in the future. Specifically, this work focuses on solving relational algebra and SQL exercises, proposing an educational solution that fosters student autonomy without providing direct answers, but rather providing support for the appropriate resolution of the problem, helping students recalibrate their reasoning to arrive at a solution compatible with the answer to the exercise. The proposal is based on the integration of two essential technological components. On the one hand, the LLM-based agentic flow is responsible for processing natural language, capturing the essence of the question and generating guiding answers, as well as implementing the reasoning flow based on the knowledge stored in the KG. On the other hand, the KG stores the theoretical knowledge derived from the course's curricular content and the pedagogical guidelines defined by the teaching staff. This structure (ontology and taxonomy) ensures that the assistance provided is consistent with the academic objectives and the particular style of the course, avoiding general LLM knowledge that might contradict the program guidelines. One of the distinctive features of the tutor agent is its ability to interact dynamically and personalized with each student. By analyzing the language and context of the query, the tutor identifies areas of difficulty and offers clues and suggestions that will guide the student in solving the exercise. This feedback process is designed to reinforce the student's progress and constructively correct errors, without specifically revealing the final answer. Furthermore, the tutor is able to link practice with theory, allowing the student to access specific theoretical details when necessary. The model incorporates the possibility of being adjusted and customized according to the needs of the course. Faculty can define and update practice patterns and objectives, allowing the tutor agent to act as a reinforcement of the course style. This adaptability ensures greater congruence between automated assistance and curricular content, facilitating the incorporation of new pedagogical guidelines. The KG ontology is oriented toward reinforcement and constant feedback, allowing both individual work on specific exercises and reasoned abstraction on specific topics or sets of exercises, favoring the agent's generalization capacity. In summary, this work represents a breakthrough in the application of emerging technologies in education, combining artificial intelligence and knowledge structuring to create a tutor agent that reinforces the learning process. The synergy between the LLM-based agentic flow and the KG results in an innovative tool that enhances the development of critical and analytical skills in students.