McLauchlan, Mark

Stanley-Davy, Emily

2025-11-14T10:39:04Z

2025-11-14T10:39:04Z

2025-08

This research explores how RAG Models can improve the Scottish geographical queries of Large Language Models (LLMs). Specifically, how the main created RAG Model GeoBot, which utilises OpenAI’s LLM GPT-4, effectively incorporates geographical data from the Gazetteer for Scotland, improving OpenAI’s ability to identify local landmarks and Scottish locations. This is supported by GeoBot’s superior Cosine Similarity and ROUGE scores compared to default GPT-4 and the created comparative no code RAG model Botsonic. However, there are still weaknesses with this approach as GeoBot performs poorly in response clarity and readability, indicating that further improvement is needed to create the strongest RAG Model methodology. GeoBot also demonstrates the viability of map generation in a RAG Model, although again some improvement could still be made to this. These results highlight the potential application of GeoBot for geospatial applications. GeoBot can be accessed through this website: https://www.geos.ed.ac.uk/dev/geobot/main. Password is ‘GIS2025!’. If this does not connect the website is now offline. The full code can be seen in the Technical Report Appendices.

https://hdl.handle.net/1842/44217

http://dx.doi.org/10.7488/era/6740

en

The University of Edinburgh

LLM

GIS

AI

RAG Model

Generative AI

Map Generation

LLM geo-chat interfaces: designing a retrieval augmented generation (RAG) model to improve large language model (LLM) geographical understanding, answer spatial queries and create Scottish maps

Thesis or Dissertation

Masters

MSc Master of Science