"Fun" with #QGIS #LLM assistants π©βπ»
Looking into the SpatialAnalysisAgent plugin today: https://github.com/Teakinboyewa/SpatialAnalysisAgent
1/n
[#]gischat #gis #geospatial #spatialdatascience
=> More informations about this toot | More toots from underdarkGIS@fosstodon.org
Compared to the two plugins I discussed in October (https://anitagraser.com/2024/10/06/llm-based-spatial-analysis-assistants-for-qgis/), this plugin looks more mature and it's publicly available for testing.
All you need is an OpenAI API key.
(Support for other / local models is planned: https://github.com/Teakinboyewa/SpatialAnalysisAgent/issues/15)
2/n
=> More informations about this toot | More toots from underdarkGIS@fosstodon.org
First experiment: using the QGIS world layer, ask the assistant to select Austria π¦πΉ
The agent π€ rumbles along and (after fixing an issue https://github.com/Teakinboyewa/SpatialAnalysisAgent/issues/21) creates a workflow diagram and some code that runs ... but doesn't provide any output ...
3/n
=> View attached media | View attached media | View attached media
=> More informations about this toot | More toots from underdarkGIS@fosstodon.org
Turns out that the plugin is having issues with Geopackage inputs (https://github.com/Teakinboyewa/SpatialAnalysisAgent/issues/19)
So, let's try again with our trusty old @shapefile
Success π€©
4/n
=> More informations about this toot | More toots from underdarkGIS@fosstodon.org
Let's take it one step further: I asked "Select all countries that border Austria"
The resulting output? -- All countries π€
It used the "Select by location tool" instead of the Extract algorithm.
Changing the prompt to "Extract all countries that border Austria" didn't help. No result this time.
The plugin invents an OUTPUT parameter that doesn't exist in Select algorithms.
5/n
=> View attached media | View attached media
=> More informations about this toot | More toots from underdarkGIS@fosstodon.org
The plugin allows users to select different OpenAI models. Based on my limited experience, gpt-4o vastly outperforms gpt-4o-mini.
Every request to the assistant creates multiple API requests, depending on different factors that I don't fully understand yet. Therefore, it is not easy to predict the costs of a single experiment.
In total my ~10 experiments resulted in 63 GPT-4o and 10 GPT-4o-mini requests, totaling $0.36 for input and output.
6/n
=> View attached media | View attached media
=> More informations about this toot | More toots from underdarkGIS@fosstodon.org
tl;dr: Don't fire your GIS analyst just yet
Maybe I'll try out the Kue plugin tomorrow. Stay tuned.
7/7
[#]GISChat #QGIS #GIS #LLMs #GeoAI
=> More informations about this toot | More toots from underdarkGIS@fosstodon.org
Really appreciate you running through this wall
=> More informations about this toot | More toots from mike.teczno.com@bsky.brid.gy
@mike.teczno.com thanks, Mike. I'd love to read what everyone else's experience with (geo) data analysis assistants has been
=> More informations about this toot | More toots from underdarkGIS@fosstodon.org
@underdarkGIS Don't fire your GIS analyst ever
=> More informations about this toot | More toots from geolaw@aus.social This content has been proxied by September (ba2dc).Proxy Information
text/gemini