Cost Minimization in Generative Dialogue Systems

This thesis focuses on minimizing the operational costs of generative dialogue systems that combine the retrieval of relevant text passages from documents or databases with response generation using GPT. This approach, called Retrieval-Augmented Generation (RAG), merges efficient information retrieval with generative techniques to produce accurate, contextually rich responses. The primary goal was to reduce operational costs while improving system accuracy. The research comprised three experiments to enhance data organization, refine structuring methods, and optimize the interaction between retrieval and generative components. A key innovation was the creation of a conditional text selection mechanism, developed based on experimental insights. This mechanism effectively reduced the generator's workload, eliminated redundant information, and improved response quality. The results demonstrated that improved data organization, thematic clustering of queries, and filtering of irrelevant information boosted response accuracy to 90% while reducing operational costs by up to eightfold. The system’s practical implementation was showcased on the Streamlit platform. This thesis offers valuable insights into designing and implementing adaptive, efficient, and sustainable modern dialogue systems. Its findings and recommendations are relevant to both academic research and commercial applications.