LLM-based RecSys: Multi-agent Architectures and Advanced RAG Techniques

This thesis presents the work carried out during my internship at Statwolf Data Science (Padua, Italy) and centers on the design and implementation of a movie recommendation system built from scratch. The system leverages state-of-the-art Large Language Model (LLM) techniques to enhance both functionality and performance, with a focus on advanced RAG strategies and agentic, multi-agent system architectures. The study begins with a quick review of Large Language Models (LLMs), deep learning foundation models based on the Transformer architecture that, in recent years, have revolutionized natural language processing and generative AI fields. These models have drastically transformed human–computer interaction through fluent dialogue, sophisticated language understanding, decision-making, reasoning, and planning capabilities, enabling the development of powerful chatbots, agents, and multi-agent systems. Foundational concepts such as transfer learning and classical Retrieval-Augmented Generation (RAG) are also introduced to establish the theoretical basis on which the rest of the thesis is built. Building on this foundation, the thesis explores three main research areas: (i) advanced RAG techniques, (ii) agent and multi-agent systems including architectures, patterns, frameworks, and (iii) the integration of LLMs within modern recommender-system pipelines. In particular, the analysis highlights how and where LLMs can be integrated into recommendation workflows. By orchestrating multiple agents and coupling them with a robust knowledge base, the system extends the capabilities of a standalone LLM to automate, coordinate, and enrich the recommendation process. The final part of the thesis details the project implementation, illustrating how the theoretical components were combined to build an enhanced movie recommendation system, and discussing the main technical challenges encountered. Overall, this research demonstrates that layering advanced RAG methods and multi-agent architectures on top of complex applications such as recommender systems can significantly improve performance, extend functionality, and enrich the user interaction experience.