Does it run and produce a trustworthy spatial result?
Check runtime status, artifact existence, spatial logic, and quality metrics together.
Project note · 2025
Verifiable GIS Code Assistant
GeoAgent
Turning GIS code generation from a one-shot answer into a loop of document retrieval, generation, execution, repair, and evaluation.
Why I started
General models often produce GIS code that looks reasonable but does not run. A wrong API, parameter, projection, or file can make an answer unusable. Users need less documentation hunting and debugging—not merely attractive code.
The real question
How can a natural-language request move through reliable knowledge, data selection, and execution feedback to become a verifiable GIS result?
Choices I made
Every action should have a reason.
Build 6k+ GIS QA pairs and 320 Query-Code examples first.
A vertical task needs stable data boundaries and example coverage; it cannot rely only on model memory or scattered prompts.
Use RAG to retrieve official documentation and code examples.
GIS APIs change and parameters are specific. Retrieval provides traceable grounding and is more maintainable than an ever-longer prompt.
Connect execution, error, repair, and evaluation in a loop.
Early review showed that code resembling an answer is not the same as a deliverable task. Execution feedback exposes dependencies, paths, tool use, and exception handling.
How the system moves
How a GIS request becomes a runnable, verifiable result
Start / inputProcessDecisionOutcome
- User
Describe a spatial task
State the goal, input data, spatial relationships, and expected artifact.
- RAG + data layer
Retrieve knowledge and confirm data
Recall official APIs and examples, then inspect files, projections, and fields.
- Planning / generation agent
Plan and generate code
Choose tools, sequence, and parameters, then produce an executable script.
Deliver code and spatial artifacts
Preserve data, parameters, and evaluation evidence for reproducibility.
Route error to the repair agent
Feed dependency, path, API, or parameter failures back into generation.
Return to the user for clarification
Request files, projection, field meaning, or a clearer expected output.
Validation loop
Route failure causes back to retrieval and planning
Locate the failure in knowledge, data, planning, execution, or evaluation, then rerun instead of hiding it in the final answer.
Business meaningThe core is not code generation alone. It is a loop that connects grounding, data selection, execution feedback, and spatial validation so plausible but undeliverable answers are exposed.
What I made
From judgment to something that runs.
Understand
Parse the spatial task, inputs, and expected artifact.
Plan
Retrieve relevant knowledge and select files, tools, and sequence.
Verify
Execute generated code, repair errors, and evaluate output quality.
What the work left behind
- Built 6k+ GIS QA pairs and 320 Query-Code examples
- 78.3% CodeBLEU improvement over baseline
- 35 GIS tasks completed in about five minutes on average
- 50% improvement in defensive coding patterns
Code similarity and a small task set do not establish production reliability. The first site release does not link the research repository because it contains local path traces.
Expand spatial-result validation and failure coverage, prioritizing “is the result trustworthy?” over “was code generated?”