Urban site selection is a complex task requiring the integration of diverse data sources and the satisfaction of multiple competing constraints, which traditionally relies on tedious GIS processing and domain expertise. This paper introduces Urban Fusion, an LLM-driven agent framework that translates natural language site selection queries into executable Python code for constraint-based spatial analysis. The framework enables users to specify complex combinations of geospatial, economic, demographic, and operational constraints when searching for optimal commercial locations.

The approach is evaluated on a dataset of Cambridge, MA parcels with over 120 site selection scenarios across varying complexity levels. Through comprehensive testing of multiple methodologies, it was found that finetuned models with chain-of-thought prompting and retrieval-augmented generation significantly outperform zero-shot and few-shot approaches, achieving a 50% success rate and 61.47% F1 score for constraint satisfaction. Error analysis reveals key challenges in spatial code generation, including schema understanding, domain-specific knowledge integration, and geospatial operation implementation. Urban Fusion demonstrates a promising approach to automating multi-constraint site selection, with implications for urban planning, commercial development, and location intelligence applications.