Viabilidade de arquitetura RAG para a geração de código em linguagens proprietárias

Abstract

The advancement of Large Language Models (LLMs) has transformed software engineering by providing support for code generation. However, the effectiveness of these tools is significantly limited in contexts involving proprietary languages due to the scarcity of public training data. This study investigates a potential solution for code generation in the proprietary language Procscript, used in the Uniface low-code platform. To achieve this, it proposes the implementation of an architecture based on retrieval-augmented generation (RAG), integrated with the Llama 3.1 8B model. The results, obtained through experiments with developers and qualitative analyses of their perceptions and the generated code, indicate that the tool is perceived as practical and useful, especially for low-complexity tasks. It was also observed that the strategy adopted by the developer while using the tool directly influences the results, with certain approaches tending to generate more effective solutions. Although challenges related to hallucinations and syntactic inconsistencies remain, the study concludes that using RAG is a viable strategy to mitigate data scarcity in proprietary languages while simultaneously meeting information privacy requirements.