Geração de sentenças judiciais utilizando LLMs e técnicas de RAG

Abstract

This study presents an in-depth analysis of the contribution of a RAG-based method for generating legal rulings confined to the Brazilian civil law domain. Drafting judicial decisions is a laborious task that demands extensive knowledge and interpretation. Common natural language processing techniques fall short of producing the depth required to be truly useful in a judge’s daily workflow. Experiments were conducted across multiple open-source large language models and hyperparameters at each stage of the system to identify an optimal configuration. The results clearly demonstrate the advantage of RAG: grounding the model’s generation in the relevant corpus of laws associated with the petition under review. This approach achieved COMET semantic parity scores of up to 73% when compared to the original decision drafted by a judge.