Urban Institute’s Minneapolis zoning benchmark had to supply the district in its prompts. Its researchers reported that models could not correctly determine the district from an address, even with web search: the interactive map was easier for people to use than machines.[1]
The missing information has several layers. Minneapolis assigns every property both a primary district and a built-form overlay; some properties carry additional overlays. The first governs uses, while built-form rules cover matters such as height, floor area and setbacks. The city provides an interactive map for identifying the applicable districts.[2][3]
For their evaluation, published in March 2026, researchers devised professional-developer and homeowner questions against a 467-page code. Retrieval supplied selected passages to models; a zoning expert supplied reference answers. Narrow searches often found accurate text but missed connected provisions elsewhere.[1]
Vague questions could produce instructions to meet yard requirements without stating those requirements. More specific prompts worked better. The benchmark therefore exposed a practical dependency: a resident may need procedural knowledge before an assistant can interpret the rules usefully.[1]
A public zoning assistant needs to help establish the property’s applicable districts as well as explain the ordinance. Testing only the explanation starts after a consequential part of the resident’s task.[1][2]
Sources & further reading
- Could Retrieval-Augmented Generation with Large Language Models Help Make Local Zoning Codes Easier to Navigate? Ridhi Purohit and Judah Axelrod / Urban Institute · 19 Mar 2026 · Accessed 29 Oct 2026
- Zoning districts City of Minneapolis · 16 Jul 2025 · Accessed 29 Oct 2026
- Zoning maps City of Minneapolis · 1 Jul 2025 · Accessed 29 Oct 2026