Real-time AI for Cities: New Architecture Brings Live Context to Urban AI Systems

Generative AI and large language models hold promise for urban applications, but they struggle with a fundamental problem: their training data becomes stale. New research proposes a framework that could overcome this limitation by continuously feeding live, location-specific information into AI systems operating in cities.

The Core Challenge#

Foundation models powering today's AI are built from data captured at a specific point in time. Once trained, they cannot easily absorb new information without costly retraining. Traditional solutions like fine-tuning work poorly in dynamic settings where conditions change by the hour—such as traffic patterns, event schedules, or resource availability in urban areas.

Retrieval Augmented Generation (RAG) has emerged as a more practical workaround: rather than updating the model itself, RAG systems feed the model fresh information retrieved from external sources at the moment a query arrives. However, most RAG implementations rely on generic semantic databases or web searches, which do not account for the spatial and temporal nature of city data.

A New Spatial-Aware Approach#

The proposed architecture integrates two key innovations:

• Real-time data injection: The system continuously pulls current information and filters it based on time and location, ensuring answers reflect the city's actual state.
• Linked data infrastructure: Implementation using FIWARE—a smart city software platform—allows seamless integration with digital twins and IoT systems already deployed in urban environments.

This combination enables the AI to answer context-aware questions that depend on knowing where something is happening and when.

Proof of Concept: Tourism in Madrid#

Researchers validated the approach through a tourism assistant for Madrid. The system can handle queries that require understanding both spatial relationships (nearby attractions) and real-time conditions (opening hours, events happening today). The framework successfully demonstrated how a foundation model can provide useful guidance that remains accurate despite the city's constant changes.

The work suggests a path forward for scaling generative AI into smart city infrastructure without the overhead of model retraining, potentially enabling more responsive and contextually informed AI services across urban systems.