The SF Planning code is complicated. Comprising thousands of pages, this document describes in excruciating detail the rules that must be followed if you dare to build something in San Francisco. Just the sheer mass of this planning code is a drag on building efficiency — the planning department itself spends significant time explaining to prospective builders how to interpret the code for their cases.
But it doesn’t have to be this way! LLMs like ChatGPT are great at ingesting dry technical language and returning it as ordinary speech. Pair this with a technique called chunking, which provides a LLM with key snippets of a larger document when answering a user’s question, and you can really start to improve the way we access this information.
PlanningGPT is an early attempt at this. PlanningGPT will do its best to answer your questions about the planning code, and it will actually cite its sources in the planning code so you can check its work afterwards.
This prototype is only scratching the surface of what’s possible here — for instance, we could equip the LLM with the ability to search for an address and get all of the zoning information relevant there. We could give it access to state housing laws, building codes, and all sorts of other information.
Enjoy, and please let me know if you have any feedback!
Thanks to Salim Damerdji and Jacob Marshal for helping me build this prototype.