USING GENERATIVE AI FOR CLASSIFICATION OF LEGAL DOCUMENTS

Resumo

The Brazilian judicial system is currently overwhelmed by an enormous backlog of digital lawsuits, making manual case sorting both financially draining and unreliable. This research explores the integration of Generative Artificial Intelligence to streamline the categorization of legal petitions through Large Language Models (LLMs). The study outlines a technical progression divided into three distinct phases. First, a few-shot learning model was tested, resulting in a modest accuracy rate of 56%. Second, the methodology was improved using prompt engineering combined with N-gram analysis and data augmentation strategies to address the issue of skewed datasets. Finally, the research implemented a Retrieval-Augmented Generation (RAG) framework to optimize performance. Using real-world data from the Court of Justice of Tocantins, the experiments demonstrated that the RAG-based system achieved a significant 84% accuracy across 11 complex legal categories. This advanced architecture effectively minimized the occurrence of AI hallucinations and clarified semantic uncertainties often found in legal texts. The findings suggest that this innovative approach provides a reliable and scalable framework for the LegalTech industry, offering a viable path toward modernizing judicial administration. By automating the initial stages of case management, the proposed solution not only enhances operational efficiency but also ensures a higher degree of consistency in the processing of legal documents, ultimately contributing to a more agile and responsive justice system in Brazil and potentially other jurisdictions facing similar digital challenges.