LLMs lack intrinsic geospatial awareness and do not maintain structured knowledge of geographic features, spatial relationships, or domain-specific dependencies. As a result, they cannot reliably reason about how disruptions propagate through interconnected spatial systems, such as transportation networks, health service accessibility, school catchments, or infrastructure dependencies during floods, wildfires, and other natural hazards. This limitation is especially important for disaster response and resilience planning, where decision-makers need to understand not only where an event is occurring, but also which people, services, facilities, and infrastructure systems may be affected. While LLMs are powerful tools for synthesizing text, they require external geospatial structures to reason over spatial networks, semantic relationships, and aggregate spatial data.

In this hands-on workshop, participants will build a geospatially aware LLM application from the ground up using open-source tools. They will create a Spatial Knowledge Graph (SKG) using Apache Jena to model networks of geographic features and their relationships, then use it to develop a Geo-GraphRAG pipeline that enables an LLM agent to translate natural-language questions into GeoSPARQL queries. This approach provides transparent, traceable answers grounded in real geospatial data, making it well suited for disaster scenarios where users need to ask questions such as which facilities are downstream of a flood event, which communities may lose access to critical services, or which transportation links are most important for response operations.

Participants will also learn how Discrete Global Grid Systems (DGGS) enable the integration of aggregate and statistical data across domains by aligning datasets to a common spatial reference. Using open-source DGGS tooling such as DGGAL, participants will build an LLM agent capable of generating both GeoSPARQL queries over feature networks and DGGS API calls using CQL2. The workshop will explore how semantic feature networks relate to DGGS zones, and how these two approaches can be combined to support richer forms of geospatial reasoning for resilience analysis, including the integration of hazard exposure, population statistics, infrastructure vulnerability, and service accessibility.

This workshop builds on our FOSS4G North America 2025 session by adding DGGS-based methods for integrating aggregate and statistical data. It demonstrates how complementary RAG structures are suited to different kinds of geospatial reasoning: graph-based RAG for reasoning over networks of features and relationships, and DGGS-based approaches for reasoning over gridded, aggregate, and statistical data. By the end of the session, participants will understand how spatial knowledge graphs and grid systems can work together to bring geospatial reasoning to AI systems, with direct relevance to disaster response, resilience, and risk-informed planning.