Towards language-based retrieval of complex geospatial data: A case study on UK National geographic database

Abstract

Retrieving granular information from large-scale geospatial databases is challenging and can require significant domain expertise, owing to complex data taxonomies and hierarchical data structures related to geometry and topology. In this research, a multi-agent system is developed that uses large language models (LLMs) for geospatial information retrieval based on natural language queries. The proposed solution employs LLMs within three agents that decompose data retrieval into sub-tasks: human instruction comprehension; data collection identification; and feature category screening. The three agents collaborate in a sequential workflow to accumulate knowledge and identify a pathway to the correct information within a hierarchical data structure. A data retrieval module subsequently interprets the search extent and retrieves data through API calls to the underlying databases. The proposed approach is evaluated in a case study with a large and complex geospatial database, the National Geographic Database operated by Ordnance Survey in the UK. Evaluation was based on a total of 83 sample queries, i.e. 40 human queries collected from volunteers, 40 synthetic queries generated by LLM, and 3 queries selected from a benchmark database. The proposed system answered majority of the queries successfully, namely 22 out of the 31 valid human queries, 29 out of 37 effective synthetic queries, and all 3 queries from the benchmark dataset. This demonstrates good potential for flexible data retrieval, locating a variety of geographical features from different data collections in response to natural language questions about tennis courts, public car parks, and restaurants, etc. The multi-agent solution was also evaluated using OpenStreetMap and the results demonstrate the transferability of the proposed method in handling different data taxonomies and structures. Our findings also highlight several research challenges around the “AI readiness” of documentation, complex geospatial data taxonomies, ambiguity in human language, and complex spatial reasoning. Overall, our study addresses two major research gaps: (1) the use of unstructured natural language to accurately retrieve geospatial information; and (2) the use of LLMs to navigate multiple heterogeneous datasets in a hierarchical data taxonomy.