Skip to project
Roman Jahandideh ← All projects
Independent project · grounded question answering

Zoning Literacy Assistant

Regulatory literacy · retrieval, synthesis & independent verification · Claude

The Zoning Literacy Assistant is designed for people who need to understand regulation but do not already speak the language of planning: homeowners, renters, neighbours, and community members as well as professionals. It investigates whether a conversational system can make zoning information accessible without reproducing the confident guessing of a general-purpose chatbot. The central design requirement is simple: every answer must show what supports it, and the system must say when the supplied document does not contain the answer.

Structured Knowledge Base

The system uses the current base provisions of Vancouver's R1-1 district as ten individually addressable clauses covering permitted uses, lot-area and frontage thresholds, setbacks, height, storeys, floor space ratio, dimensions, outdoor space, parking, and tree retention. Structuring the bylaw into retrievable clauses makes the evidence behind an answer visible and allows the system to reason about provisions that interact with one another.

Retrieve, Answer, Verify

Each question passes through three independent stages. Retrieval receives the question and clause titles and selects the relevant provisions. Synthesis receives only the selected clause text and returns a plain-language answer, a grounded true-or-false flag, and exact supporting quotations. Verification then checks whether the answer is supported by its own evidence and whether an omitted clause should have been retrieved. When retrieval finds no relevant provision, the pipeline stops and reports that the question is not covered rather than allowing a second model call to fill the gap from general knowledge.

Deterministic Site Assessment

A second mode accepts lot width, lot depth, and rental or below-market housing status and computes the unit-count tier, buildable footprint, permitted floors, and estimated gross floor area through ordinary deterministic code. The language model is used only to explain those already-computed results and cite the relevant provisions. Toggling the rental condition demonstrates the real R1-1 rule by changing the floor space ratio from 0.70 to 1.00 and recalculating the downstream figures.

Evaluation and Deliberate Failure Case

The prototype includes a deliberately uncovered question about the number of storeys permitted for a laneway house. The source clause confirms that a laneway house is permitted but does not provide its height or storey limits. The correct behavior is therefore to return “not covered,” making the system's restraint observable rather than merely claimed. End-to-end tests with controlled model responses verify retrieval, citation rendering, verification, API-key handling, and deterministic feasibility calculations. For a 10 by 37 metre lot, the checked example produces a 162.6 square metre buildable footprint and shows the change from one floor at 0.70 FSR to two floors at 1.00 FSR.