LLM-enabled spatial query with examples

Abstract

Efficient spatial search is essential for modern information retrieval. Example-based approaches have emerged as alternatives to address the limitations of traditional search engines. This report introduces GeoGPT, an advanced example-based spatial search system enhanced by Large Language Models (LLMs) such as GPT-4 and Llama 2. GeoGPT overcomes the shortcomings of traditional approaches by leveraging LLMs to effectively interpret complex spatial relationships, refine nuanced user queries, and adapt to evolving user intents in real-time. To support robust user interactions, a graphbased data pipeline has been developed to generate synthetic data that closely mirrors real-world user behaviour. This enables precise fine-tuning of LLMs for spatial queries through techniques such as Parameter-Efficient Fine-Tuning (PEFT) and Instructional Tuning.

Additionally, GeoGPT features an intuitive user interface with interactive tutorials, intuitive dual-mode functionality, and a user-friendly map. This study lays the foundation for enhancing user experiences and expanding LLM applications in information retrieval across diverse domains.