Axis Agents
← Blog

16 November 2026 · Build story

How we built an AI chat for a spatial organisation

The AI chat for spatial organisations connects one chat to the documents, spatial data and tools an organisation chooses. Staff ask a question and open the map or file that comes back, with the sources it used. The case study shows the interface and a worked result. This post covers how we built it.

Spatial teams switch tools to answer one question

A spatial team keeps its knowledge in documents, catalogues, data stores and specialist software. To answer one question, someone moves between several of them and carries the context by hand.

We wanted one conversation instead. The place, the date range and the source stay attached to the question, so a colleague can follow any answer back to where it came from.

We started from an open interface

We did not build a new chat window. The interface is the open-source LibreChat, which Axis Spatial specialised for spatial work. People already know how to use a conversation. The work went into what sits behind it: the connections, the execution, the saved work and the checks.

Three steps for the user

  • Choose what to connect: the documents, catalogues, spatial data, APIs and specialist software the chat may use, and nothing else.
  • Ask a spatial question in a normal conversation. The system keeps the place, the date and the source with the task.
  • Open the work: follow the source record and examine the map, raster, vector output or interactive file the connected tools returned.

A fluent answer can lose the place

A place name, a date range and a coordinate system decide what an output means. A chat that drops them gives an answer that reads well and describes the wrong thing. It reads coordinates in the wrong system, and the map still looks fine.

So we carry the place, the date, the source and the coordinate system through the workflow, and we check what each output represents. Spatial code and tools run in an isolated environment with agreed limits. Sessions, files and checkpoints are saved, so people can inspect the work and continue it later.

A worked example: Helsinki

The demo uses a real Sentinel-2 observation of Helsinki from 27 June 2024. The place, a date range and the Sentinel-2 source go in. An isolated raster workflow with bounded output calculates a vegetation index (NDVI). The output is a Cloud Optimised GeoTIFF, its metadata and a colour-rendered map, linked to the public source record.

The map uses red for low values and green for higher ones. The case study page changes only how the map is drawn. It does not change the calculation.

We choose the model with the Spatial Intelligence Index

A language model runs each job, and the choice matters. We choose the model for each job using the Spatial Intelligence Index, our sister lab's public benchmark at axisspatial.com/arena. It compares models on spatial tasks, so we pick on those results and not on general reputation.

A person reviews the interpretation

In one review of the hosted demo, the first answer misexplained how water behaves in the imagery and overstated the width of the area. A follow-up corrected the explanation without recalculating the map. That is why a person reviews the interpretation. We agree the checks and review points with each team and build them into the work.

The hosted demo is invitation-only. An operator checked it on 13 September 2026: the owner login, a real STAC tool result, source opening and rendered-map opening all worked through the stock LibreChat interface. The chat screenshot on the case study is from April 2026 and is historical.

Building one for your organisation

For a real deployment we set up the access, the connections, the compute, the saved work and the review process around your team and the job. Send us one task your team does today, with the data it uses and the software it runs in. We will say which model can do it and what it would take. Book a call to start.