Technical Summary: Real-time Spatial Retrieval Augmented Generation for Urban Environments

Problem Statement

The proliferation of Generative AI and Large Language Models (LLMs) offers transformative potential for urban applications through "Urban Foundation Models" (UFMs). However, base models are limited by their static training data, making them unsuitable for dynamic urban environments characterized by real-time data updates, interconnected systems, and massive data volumes. Traditional methods like fine-tuning are too costly and slow for frequent updates, while standard Retrieval Augmented Generation (RAG) architectures—relying on semantic databases, knowledge graphs, or web searches—fail to meet the specific demands of urban contexts. These traditional RAG systems often lack the ability to perform real-time spatial filtering, handle high-frequency updates from IoT sensors, and manage the complex relationships between physical urban elements without introducing excessive latency or hallucinations.

Methodology

The authors propose a Real-time Spatial RAG architecture designed specifically for urban environments, implemented using the FIWARE ecosystem. The methodology integrates the following components:

1. Architecture Design: The system leverages FIWARE's NGSI-LD (Next Generation Service Interface for Linked Data) standard to manage context information. It employs a Context Broker (OrionLD) to handle real-time data ingestion from IoT sensors and external sources.
2. Spatial and Temporal Filtering: Unlike naive RAGs that retrieve based solely on semantic similarity, this architecture utilizes the Context Broker to perform preliminary spatial filtering (based on geospatial location) and temporal filtering (based on real-time updates) before data reaches the LLM. This reduces the context load on the model.
3. Data Flow:
   • Update Phase: IoT devices or external APIs update entities in the Context Broker.
   • Retrieval Phase: When a user query is received, the system queries the spatial database (Context Broker) using the user's specified region and filters. It retrieves relevant NGSI-LD entities, which can be further refined via semantic retrieval and reranking.
   • Generation Phase: The filtered entities (in NGSI-LD format) and the user prompt are sent to the UFM. The model is instructed to answer based only on the provided retrieved data.
4. Implementation: The architecture was implemented using FIWARE components (OrionLD, Draco, IoT Agents) and tested with various LLMs (including GPT-5.2, GPT-4.1, Llama 3.1, Qwen, and DeepSeek).

Key Contributions

• Novel Architecture: The paper introduces a reference architecture for Real-time Spatial RAG, distinguishing it from Naive RAG, Graph RAG, and Web-based RAG by explicitly incorporating geospatial filtering and real-time data handling via Linked Data standards.
• FIWARE Integration: It demonstrates the feasibility of integrating Foundation Models into existing smart city ecosystems by extending the FIWARE digital twin architecture, ensuring interoperability, security, and scalability.
• Empirical Validation: The study validates the architecture through two distinct use cases in Madrid:
  1. A Tourism Assistant providing real-time recommendations for Points of Interest (PoIs) based on occupancy, price, and location.
  2. A Traffic and Lighting Management System monitoring the status of 650 traffic lights and luminaires, including crowd density and operational states.
• Performance Analysis: The authors provide a comprehensive evaluation of correctness, precision, recall, and latency across different model sizes (commercial and open-weight) and data volumes (Small: 10–20 entities, Medium: 100–200, Large: 650–1300).

Results

• Correctness and Hallucination: As the volume of retrieved entities increases (moving from Small to Large regions), the proportion of "OK" (correct) answers generally decreases across all models due to the difficulty of processing large contexts without hallucinating.
  • GPT-5.2 demonstrated the most consistent performance and best hallucination control in large-scale scenarios, prioritizing safety over recall.
  • GPT-4.1-mini excelled in small and medium contexts but suffered from increased hallucinations in large scenarios.
  • Open-weight models (e.g., DeepSeek-V3-685B, Llama-3.1-70B) showed competitive performance, though smaller models (8B/7B) failed in large scenarios due to context window limitations.
• Ablation Study: Comparing different RAG configurations revealed that Spatial RAG significantly improves precision and recall in small and medium regions by filtering irrelevant data. However, for large regions, combining Spatial RAG with Semantic RAG and reranking is necessary to balance the number of entities passed to the LLM with retrieval accuracy.
• Latency: The primary bottleneck is the LLM inference time, which scales linearly with the amount of context. The Context Broker and retrieval components contribute negligible latency.
  • LLM response times ranged from ~1.1 seconds (small context) to ~2 seconds (large context), with significant variability.
  • The authors note that current LLM latencies make hard real-time applications (e.g., autonomous driving control loops) impractical; the system is currently limited to soft real-time scenarios (updates in seconds).
• Cost: There is a significant trade-off between model performance and cost, with larger open-weight models offering economic advantages for large-scale deployments despite higher inference costs per token for some commercial models.

Significance and Claims

The paper claims that its proposed Real-time Spatial RAG architecture successfully bridges the gap between static Foundation Models and the dynamic requirements of smart cities. By offloading spatial and temporal filtering to the Context Broker (FIWARE) prior to generation, the system effectively manages the scalability and latency issues associated with processing massive urban datasets.

The authors emphasize that while the integration of UFMs into cities is feasible for soft real-time applications (like tourism assistants and traffic monitoring), significant challenges remain regarding hard real-time constraints and the tendency of models to hallucinate when processing large volumes of dynamic data. They conclude that the architecture provides a robust, interoperable foundation for future urban AI, but future work must address scalability to millions of entities, improve model efficiency for edge deployment, and explore agentic capabilities for multi-hop reasoning. The work does not claim to solve all urban AI challenges but establishes a validated reference architecture for integrating generative AI into existing smart city infrastructures.