PoliRural+ — Bridging Rural and Urban Europe · knowledge baseInteractive book · llms.txt · For AI agents

GeoAI: grounding language models in place

A chatbot guesses about little-known places. GeoAI makes it fetch data for the exact area on the map before it answers.

What does GeoAI add to a general-purpose language model, as PoliRural+ designs it?

It ties each question to an area on a map and enriches it with retrieved spatial data, metadata and tool results before the model answers. The model interprets and explains, while specialised services do the data access and calculations, so the answer holds for that place, scale and time.

Open in the interactive book · missions, mini-games and memory cards there

What GeoAI adds: a question enriched with geodata before the language model answers A language model alone gives generic answers about lesser-known places. GeoAI, as JackDaw implements it, parses the question and the area drawn on the map, fetches matching open geodata and project documents for that area, and adds them to the prompt. The model then returns an answer tied to the area, with tables or maps and references to the datasets used. WITHOUT GEOAI Same question Language model only Generic text: LLMs struggle with lesser-known places (D4.1 §6.3) WITH GEOAI (JACKDAW) Open geodata layers and project documents Copernicus INSPIRE OpenStreetMap Eurostat · GISCO LPIS · cadastre policy documents Question + area natural-language question, plus an area drawn on the map GeoAI enrichment · parse topic, place, time · find matching datasets · fetch data for the area · add it to the prompt Language model reasons over the added evidence; GPT (OpenAI) or self-hosted Ollama, Llama CPP Located, sourced answer · text tied to the area · tables, charts, maps · references to the datasets and documents Takeaway: the model answers from the area's own data, so each figure can point back to its source. Sources: D4.1 §2.2, §5.2.4, §6.3 · D4.2 §2.3, §3.3
A language model alone gives generic answers about lesser-known places. GeoAI, as JackDaw implements it, parses the question and the area drawn on the map, fetches matching open geodata and project documents for that area, and adds them to the prompt. The model then returns an answer tied to the area, with tables or maps and references to the datasets used.

What is GeoAI? A language model that looks up the place first

main tellingdepth: standard3 min markdown

In PoliRural+, GeoAI means a language model that does not answer a question about a place until it has looked the place up. The platform design (D4.1) states the problem it solves: large language models struggle to answer accurately about lesser-known locations or areas with limited information on the internet. For a small rural municipality, that information may be very thin.

Enrich the question, then answer

D4.1’s remedy is to enrich the user’s query with spatial data before it reaches the model. The user asks a question in plain language and marks an area on a map. The system identifies the topics and the spatial and temporal resolution the question needs, searches catalogues for suitable data sources, services and models, calls them, and adds the results to the model’s context. Only then does the model write its answer, alongside tables, charts or maps of the retrieved data.

The Beta platform report (D4.2) describes the division of labour: the language model works as a controlled orchestrator of verifiable spatial analysis rather than an isolated text generator. It interprets, coordinates and explains; authoritative calculations and data access stay in specialised services. Because only the relevant datasets and tools are selected, the model is also spared the context overload of reading everything at once.

What changes in the answer

D7.2’s practice abstract GeoAI meets LLMs compares pure LLM output with JackDaw’s tool-calling approach on three questions:

QuestionPure LLMGeoAI agent (JackDaw)
What is the weather here?A generic explanation or a pointer to other sourcesCalls a weather API for the area on the map
Is this plot suitable for potatoes?Broad advice on potato farmingCombines land cover, elevation and weather for that plot
Which water bodies are here?May cite the wrong river or a distant cityInspects land-cover layers in the area

The gain does not come from a cleverer model. It comes from a different source of facts: the answer is checked against data for a specific place, scale and time period, which D4.2 names as the condition a territorial answer must meet.

The raw material

D4.1 lists the data GeoAI draws on: Copernicus Earth observation, INSPIRE, Eurostat and GISCO, SoilGrids, Natura 2000, Open-Meteo, FADN and OpenStreetMap, plus the project’s own Open Land Use database and local maps and IoT data. How satellite imagery actually enters is the subject of Copernicus in JackDaw.

GeoAI is the idea. JackDaw is how PoliRural+ builds it, released as a Beta in D4.2 within the project’s wider aims (PoliRural+).

Related

Sources