Agentic and Privacy-Preserving Retrieval-Augmented Generation: Architecture and Empirical Evaluation

In recent years, large language models (LLMs) have demonstrated strong capabilities in natural language understanding and generation. Trained primarily on publicly available data, these models are unable to answer queries related to private and sensitive organizational data, as they lack access to private knowledge sources. To overcome this limitation, two main approaches are used: fine-tuning models for specific domains or employing Retrieval-Augmented Generation (RAG) to retrieve relevant information from private knowledge bases. Fine-tuning requires high computational resources and expertise, often beyond the capabilities of small and medium-sized companies, making RAG a more practical solution. This thesis investigates the use of RAG and agentic solutions to enable LLMs to operate over domain-specific data while preserving sensitive information. It also examines the design of a multi-agent system capable of autonomously executing user tasks. The proposed AI framework was developed and evaluated in collaboration with a transportation company and an Italian banking institution, involving system design, implementation, and experimental evaluation of retrieval performance and privacy preservation strategies.