Utilização de técnica RAG para geração de diálogos em RPGS

Abstract

This paper investigates the use of Large Language Models (LLMs) and the Retrieval-Augmented Generation (RAG) technique to assist Game Masters in creating coherent and contextually accurate dialogues in Role-Playing Games (RPGs), considering the complexity imposed by multiple rulebooks and expansions. The research involved significant methodological challenges, from implementing context retrieval to experimental validation. The experiment was conducted in the RPG Ordem Paranormal, one of the most influential in the Brazilian scenario, using crowdsourcing for evaluation. Eleven players analyzed four dialogues — two written by humans and two generated by the model (with and without RAG) — across five categories: engagement, coherence, cohesion, creativity, and surprise. The results showed that while human-written dialogues surpass generated ones in creative aspects, the RAG technique significantly improves generated dialogues, outperforming those without context retrieval in engagement (27.3% vs. 9.1%), coherence (27.3% vs. 18.2%), creativity (27.3% vs. 18.2%), and surprise (36.4% vs. 18.2%). It is concluded that RAG is an effective solution for enriching narratives in tabletop RPGs, making dialogue generation more dynamic and context-aware without requiring model retraining.