01 / Research question
Can general-purpose LLMs reliably generate spatial SQL involving distance, geometry predicates, and spatial joins?
02 / Method
- 01
Construct a PostGIS benchmark pairing natural-language questions with executable spatial queries.
- 02
Evaluate state-of-the-art models with schema-aware prompting and augmented question–query pairs.
- 03
Measure behavior with Execution Accuracy and Valid Efficiency Score, then classify failure modes.
03 / Outcome
The resulting system demonstrates competitive executable retrieval while establishing a reproducible basis for studying LLM behavior on spatial databases.