This talk presents two complementary methods for introducing geospatial awareness into LLMs by shifting interoperability to the point of data publication rather than relying on downstream, point-to-point integration. It also demonstrates how LLM agents can orchestrate between these methods, selecting the appropriate geospatial reasoning structure depending on the question being asked.

Discrete Global Grid Systems (DGGS), an Open Geospatial Consortium (OGC) standard, support common geographies by quantizing data into hierarchical grid cells identified by standardized zone IDs. DGGS are especially well suited for integrating aggregate and statistical data across domains, such as hazard exposure, population estimates, environmental indicators, and infrastructure vulnerability. When datasets are published using the same DGGS reference system, they become more readily interoperable across domains. LLM agents can then translate natural-language questions into DGGS API calls with CQL2 parameters, enabling queries such as identifying areas where flood levels exceed specified thresholds or where hazard exposure intersects with vulnerable populations.

Spatial Knowledge Graphs (SKGs) provide a complementary capability by representing networks of interconnected geographic features and their semantic relationships across domains. SKGs are especially well suited for reasoning over networks of knowledge, such as relationships among infrastructure, populations, services, administrative units, facilities, and hazard-prone areas. Through Geo-GraphRAG, LLM agents can translate natural-language questions into GeoSPARQL queries, enabling transparent reasoning over feature-level relationships and dependencies.

Together, DGGS and SKGs allow LLM agents to use the best available structure for each type of geospatial reasoning task. DGGS provide a powerful approach for integrating gridded, aggregate, and statistical data, while SKGs provide the semantic feature networks needed to reason over relationships, dependencies, and propagation effects. By orchestrating between DGGS API calls and GeoSPARQL queries, an LLM agent can combine statistical spatial analysis with graph-based reasoning, supporting richer forms of decision support while preserving transparency and traceability in model outputs.