Development and Optimization of a Retrieval Augmented Generation System for Enhanced Conversational AI Assistance.

For a considerable time, chatbots have acted as helpful means to assist users in retrieving information. However, with the advancement of generative language models, it has become possible to elevate these bots; they are now able to understand user needs and respond to them. This thesis explains the creation of a platform for building a system which integrates individual approaches to the construction of chatbots, developed during the internship in TownHall Reply S.P.A. This research aims to investigate generative language networks incorporated in a chatbot system to elevate user interaction and satisfaction by presenting a more relevant and contextual response. The analysis extensively covers all the steps involved in creating the chatbot, subjecting it to tests specifically designed for this purpose. The central objective of the research is to thoroughly examine the challenges the system must face, with a particular focus on responding in a complete and accurate manner. The proposed solutions employ various methodologies, including the use of different databases, testing multiple embedding models to generate suitable vector spaces, the application of RAG and Reranking to achieve more precise results, and leveraging Large Language Models (LLMs) to formulate appropriate human-like responses, as well as the importance of prompt engineering. Each of these aspects constitutes a key discussion point, where tests are conducted to find the optimal setup, considering time, resources, and accuracy. The results contribute to a deeper understanding of potential issues in the fields of Natural Language Processing and Deep Learning, while also laying the groundwork for future advancements in this increasingly utilized domain.