Comparative Analysis of Direct Prompting and Retrieval- Augmented Generation for PL/SQL Educational Support

Virtual assistants for students represent an important application of Large Language Models in education by offering academic support without program limitations. However, finding the model that works best as an assistant and measuring the effect that Retrieval-Augmented Generation has over direct prompting in technical courses are aspects that can still be explored. We configured, tested and compared six virtual student assistants for a PL/SQL Database Programming course. We selected three Large Language Models, including OpenAI GPT-5.2, Anthropic Claude Opus 4.6 and Google Gemini 3 Pro, for analysing two different scenarios. In the first one, the models received a system prompt to guide their behaviour and the response was based only on the models’ built-in knowledge. For the second scenario, beside system prompting, we used Retrieval-Augmented Generation, where answers were anchored in the official Oracle documentation. To evaluate the results, we used the RAGAs framework, which looks at faithfulness, answer relevancy, answer correctness, context precision. For the evaluator model in RAGAs framework, we selected GPT-4o. Standard natural language processing metrics, like ROUGE, BLEU, METEOR and BERT Score, were used to round out the evaluation. The results show that the access to official documentation improves faithfulness across all models. GPT-5.2 achieved the highest faithfulness score of 0.69. Answer correctness scores show a difference between the reference answers and the longer and more detailed model responses. The study offers a framework for evaluating multiple models and gives practical advice for institutions that want to integrate virtual student assistants in technical classes.