Many organizations are working to harness LLMs for geospatial understanding, but face significant challenges including hallucinations, non-determinism, and the "black box" nature of these models. The geospatial domain introduces additional complexity, as LLMs are primarily designed for text-based tasks rather than spatial reasoning.

This talk introduces two open source libraries we've developed to address these challenges:

• https://github.com/Element84/e84-geoai-common/ - A foundation for geospatial LLM tasks that enables conversion of natural language to structured Pydantic models, supports batch processing, and includes robust evaluation frameworks.
• https://github.com/Element84/natural-language-geocoding/ - An innovative library that converts natural language descriptions like "within 2 km off the coast of Maui, west of Kahului" into precise polygons.

We'll explain why we built custom libraries instead of adapting existing frameworks. While solutions like LangChain and DSPY offer powerful features, they often require contorting your code to fit within complex, deeply nested class hierarchies. Our approach allows you to leverage LLMs while effectively containing their non-deterministic behavior, enabling natural language capabilities without restructuring your entire development approach. Our libraries specifically address the ambiguity in spatial language while maintaining the precision geospatial applications demand.

The e84-geo-ai-common library provides foundational tools for geospatial AI/ML work, with special emphasis on LLM integration. We'll demonstrate its core components:

• Base LLM classes with clean abstractions that simplify implementation details
• Implementations for multiple models with batch support
• Utilities for processing natural language into structured Pydantic models
• Enhanced geospatial utilities built on Shapely for specialized operations

The natural-language-geocoding library represents a significant innovation in converting descriptive language into geometric representations. We'll showcase how it transforms expressions like "within 2 km off the coast of Maui west of Kahului" into precise polygons through:

• Representation of spatial queries as directed computational graphs
• Support for complex spatial operations including intersection, union, coastal features, borders, and more
• A customizable geocoding database built from open data sources like Who's On First and Natural Earth
• Handling of colloquial place references and geographic hierarchies

Throughout the presentation, we'll emphasize our design philosophy centered on reliability. We'll share concrete examples of how we've addressed challenges through:

• Strategic design choices that focus LLMs on their strengths (translating natural language to spatial operations) while using traditional geospatial libraries for geometric precision
• Robust error handling patterns to manage edge cases in spatial descriptions
• Comprehensive evaluation frameworks for measuring and improving accuracy