Desvendando a poesia com IA: influência do ajuste de hiperparâmetros no RAG para a compreensão de texto poéticos.

Abstract

This study explores the application of the Retrieval-Augmented Generation (RAG) system to poetic texts, focusing on the customization of its hyperparameters to optimize understanding and textual generation in a literary genre that is challenging due to its semantic and structural density. Within the context of Large Language Models (LLMs), RAG presents itself as a valuable tool to overcome the limitations of fixed knowledge, dynamically integrating updated information from external sources. This work employs a quantitative methodology to evaluate the effectiveness of RAG, using the Correctness metric to measure performance and manual analyses to refine the results obtained automatically. By modifying hyperparameters such as chunk size, chunk overlap, and generation model, the study aims to determine the ideal configuration for generating precise and relevant responses to questions about poetry. The findings reveal that precise adjustments to these parameters influence the quality of the information retrieved and the responses generated, highlighting RAG's ability to produce enriched and contextually appropriate answers.