HARMONIA

To ensure reliability and inclusiveness, HARMONIA combines cutting-edge AI techniques with structured, context-aware knowledge integration. Our approach enhances Large Language Models (LLMs) using two complementary strategies: Retrieval-Augmented Generation (RAG) and Knowledge Injection. These methods allow models to generate responses that are informed by real-world, up-to-date, and verifiable information—reducing hallucinations and improving relevance.

A central element of trustworthiness lies in data transparency and fairness. We are developing a framework for data collection and encoding that is modular, scalable, and transferable to other cities.

The system is designed to reflect different ways people perceive social issues, using what we call a perspectivist approach. This means the model doesn’t just answer based on majority trends, but can generate tailored responses aligned with the needs and viewpoints of specific citizen groups.

By integrating these techniques with behavioral theories of decision-making, HARMONIA aims to create LLMs that are not only powerful but also transparent, adaptable, and sensitive to the complexity of real-world governance. Our goal is to make AI a trustworthy partner for public administrations—supporting fairer, more effective decisions for all.