PropTech · RAG
Jurisdiction-aware permitting RAG
A retrieval assistant that guides US users through construction permitting rules by location — grounded answers, not generic chat.
- Pattern
- Production RAG
- Grounding
- Jurisdiction docs
- Interface
- Guided assistant
- Risk focus
- Hallucination control
Problem
Construction permitting rules vary sharply by jurisdiction. Generic LLM answers invent requirements and create legal/operational risk.
The product needed retrieval constrained to authoritative sources, clear citations, and a UX that walks users through steps — not an open-ended chatbot toy.
Approach
Document ingestion and chunking tuned for regulatory text; vector retrieval with metadata filters by jurisdiction.
Answer generation required cited passages; refusals when retrieval confidence was weak.
FastAPI backend with evaluation harnesses for groundedness and regression tests when corpora updated.
Outcome
Shipped a production RAG assistant that answers from jurisdiction-specific material with citations. Product risk shifted from ‘model invents law’ to a measurable retrieval quality problem the team could improve.
Stack
- RAG
- Vector DB
- OpenAI / Anthropic
- FastAPI
- Eval harness
- React