FOSS4G NA 2024

Geospatial Workflows with Large Language Models
09-11, 11:30–12:00 (America/Chicago), Grand G

A discussion of techniques for integrating your application directly with an LLM and how you can work through some of the common pitfalls associated with it.


When given an overwhelming amount of options, it can be challenging to help a user get to the result they are looking for. Historically we would just add one more filter option, one more search box, or one more tutorial/help feature. In today's environment, you can ask popular chatbots questions like "what type of data" or "how can I use this file", but there is still a manual process of interpreting and implementing the result. In this talk, I'll discuss how we can help an LLM get to the answer we need to feed into the very specific use cases an application.

There will be a demo of CartoBot. A python backed web application which takes unstructured user input, like “zoom in”, “fly to paris”, or “color the buildings purple” and leverages ChatGPT to manipulate the map for the user. I discuss some of the techniques used to make this work in a web map. Additionally I'll provide an example of how you could use the same techniques for data selection in an open-source application, Eventkit-Cloud.