A permits chatbot must stop inventing ordinance language and instead cite city code. What prompt-design practice helps most?
Select an answer to reveal the explanation.
Short Explanation
If the bot is going to talk ordinances, make it show its homework. Grounded prompts that force citations of municipal code keep permits answers honest instead of creative fiction.
Full Explanation
Designing grounded prompts that require citation of municipal code sections is the primary prompt-design control when a permits chatbot must stop inventing ordinance language. Grounding instructs the model to retrieve and quote city code (or RAG passages tagged with section numbers) and to refuse unsupported legal-sounding claims. Requiring explicit citations turns ordinance answers into traceable references residents and clerks can verify, reduces hallucination of section numbers and fine amounts, and aligns the bot with how municipal permitting staff already work—from published code, not improvisation.
Defining clerk persona prompts that specify expertise scope and answer style rules fails because defining scope and style does not require the model to retrieve or cite actual municipal code sections, so it can still invent ordinance language while sounding authoritative. Providing few-shot examples demonstrating desired citation style and answer structure fails because showing the model what a citation looks like does not require it to pull and verify a real section number from the actual municipal code, so answers can still be fabricated. Instructing the model to synthesize code sections into concise plain-language summaries fails because synthesizing lets the model rewrite and combine rather than cite exact code, making it easy to drop section numbers or invent ordinance language in plain words.
Exam caveat: Grounded citation prompts are necessary but not sufficient alone; they must be paired with retrieval quality, allow-lists of authoritative corpora, and evaluation of groundedness. Citation format requirements in the prompt do not replace content-safety or disclaimer policies for legal-adjacent civic assistants.
Operational check: Run an evaluation set of permit questions where expected answers include known code section IDs; score groundedness and citation accuracy; confirm the prompt template requires section references or an explicit inability-to-cite refusal; spot-check production transcripts for unsourced ordinance claims; and ensure the retrieval index actually contains the current municipal code versions clerks use.
Prefer to study offline? Printable PDF study book + test engine ($7.50, unofficial). These questions stay free here.