Prompt to Path: LLM-Guided Multi-Objective Eco-Routing via Geohash-Compressed Urban Graphs

Cycling is a widely-practiced, eco-friendly mode of transportation essential for sustainable urban mobility. Determining routes that optimize for environmental factors, such as better air quality (AQ) and high vegetation (NDVI) as opposed to just faster routes is an important first step in incentivizing bike riding as a form of transportation. However, existing navigation systems often fall short and fail to facilitate this multi-objective, environment aware routing. To address this, we introduce a scalable eco-routing framework Geohash-Based Multi-Objective Bike Routing (GB-MOBR), that leverages Large Language Models (LLMs) to interpret natural language prompts for route preferences. The challenge lies in the prohibitive computational cost of multi-objective algorithms with large scale graph-based networks. Therefore, we first construct a fine-grained, multi-criteria graph by associating environmental data (AQ, NDVI, Congestion, Green Roofs, Meteorological) with a street network using KD-Trees for efficient spatial matching. Then, to enable scalable optimization, we apply a Geohash-based graph aggregation technique to create a smaller, topologically equivalent graph. We compare the performance and solution quality of these algorithms: Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Dijkstra and Yen's K-Shortest Paths on both graphs and examine their performance. Results demonstrate that our aggregation method reduces computation time by up to 90% and therefore making the algorithms more viable for real-time use.