Agentic Urban Planning Assistant: A Parametric Approach to Spatial and Regulatory Evaluation in Urban Planning

Urban planning decision support often requires a joint interpretation of legal rules, parcel geometry, zoning parameters, street context, and proposal-specific constraints. Generic retrieval-augmented generation is insufficient for this setting since many answers depend on structured spatial evidence and deterministic regulatory calculations rather than on retrieved prose alone. This paper presents an agentic urban planning assistant designed for parcel-level first-pass analysis. It combines deterministic request classification, route-specific retrieval, parcel lookup, exact-index zoning-parameter resolution, parcel-side normalization, setback-rule assignment, gated planning calculations, and final large-language-model (LLM) answer synthesis. The implementation uses the LLM mainly for the final answer generation stage, while the evidence and reasoning foundation is constructed through structured data preparation and deterministic workflow logic. An evaluation framework is developed that aligns benchmark design and metric selection with the actual assistant’s route structure. It combines deterministic checks, complementary LLM-as-a-judge metrics, benchmark generation from reviewed seed cases, and execution-level observability. A reviewed 40-case architecture-aware benchmark for the Sofia/Bulgarian case study is prepared. On this benchmark, the deterministic pass rate reached 1.00, while the complementary judge-based scoring yielded a faithfulness value of 0.931 under a revised structured-evidence evaluation setup. The remaining weaknesses are concentrated in proposal-feasibility synthesis and parcel-zoning descriptive completeness.