ETD

• Chatbot
• Large Language Models
• Reinforcement Learning
• Retrieval-Augmented Generation

Enterprise technical software represents not only the operational foundation of organizations but also the specialized expertise accumulated over time. However, its maintenance and effective use present several challenges, often due to complex and outdated codebases, reliance on niche programming languages and frameworks, and the need for highly trained workers.

This study investigates the potential of Artificial Intelligence, and in particular Large Language Models, to address these challenges in the case of a legacy modeling tool developed in Fortran in the 1970s for chemical plant design and flow diagram definition.

The first phase focuses on code maintenance, exploring the specialization of an LLM for Fortran code generation. A dedicated benchmark was developed by adapting HumanEval to Fortran, revealing the limitations of current models in Fortran code generation. Through a combination of supervised fine-tuning and reinforcement learning with synthetically generated data, the model achieved a substantial improvement in benchmark performance, demonstrating the feasibility of training domain-specific code generation models with limited resources.

The second phase addresses the problem of software usage through the development of a chatbot assistant for the case study system. Implemented with a Retrieval-Augmented Generation architecture, featuring vector retrieval and a knowledge graph, the assistant was deployed on Microsoft Azure and designed to support both learning and common interactions with the application.

The results confirm that LLMs can play a meaningful role in enhancing the sustainability and effectiveness of enterprise technical software, by helping the modernization of legacy codebases and improving user interaction.