O uso da geração aumentada por recuperação na validação da extração de argumentos de documentos jurídicos

Abstract

Large Language Models (LLMs)represent a transformative technology in several areas, including the legal sector. However, the propensity of these models to generate non-factual information, known as "hallucinations,"poses a significant risk to their application in contexts that demand high precision. This work proposes and evaluates a strategy that adapts the Retrieval-Augmented Generation (RAG) architecture as a mechanism to validate arguments extracted from legal documents. The main objective is to extract arguments from defense briefs and then validate them by cross-referencing them with the most similar excerpts from the source document itself, thereby increasing the process’s reliability. The methodology included a comparative analysis between open-source and proprietary models, and the development of a system that provides explicit traceability for each generated argument. The results indicate that validation is crucial and that the proposed approach increases confidence in the tool, allowing even free models to achieve robust performance. It is concluded that the effective integration of AI in law depends on hybrid systems that combine automation with qualified human supervision.