CodesWise vs ChatGPT for Canadian Building Code Questions
ChatGPT is a remarkable general-purpose tool. It is also the wrong tool for code compliance work. Here is the honest, side-by-side breakdown of why a citation-grade reference engine is required, and what changes when you use one.
The hallucination problem
General-purpose LLMs like ChatGPT do not have a copy of the National Building Code, the National Fire Code, the National Plumbing Code, NECB, or any provincial code in their training data in any reliable way. They were trained on text scraped from the public web, which contains fragments of code commentary, forum posts, and stale excerpts, but not the canonical published text.
When you ask ChatGPT a code question, it generates an answer that sounds plausible because it has seen many similar-looking sentences. It does not retrieve the actual clause. It does not verify the page number. It will confidently cite Section 9.10.5.1.(2) when the correct reference is 9.10.9.18.(1), and the only way you would know is to open the code and check.
For drawing review, permit submission, or inspection work, that failure mode is unacceptable.
What CodesWise does differently
Every answer in CodesWise comes from a retrieval pipeline that runs against the actual published code text. We ingest the National Building Code 2020 and 2025, the National Fire Code, NPC, NECB, the Ontario Building Code 2024, BC Building Code 2024, BC Fire Code, the Vancouver Building By-law 2025, the NBC Alberta Edition, the NFC Alberta, and the Code de construction du Québec directly from the official PDFs published by the National Research Council of Canada and provincial publishers.
When you ask a question, the system:
- Searches the actual code text (not training data) using a vector similarity model
- Reranks the top candidates with a cross-encoder for precision
- Sends only the matching passages to Claude, with a system prompt that forbids unsupported claims
- Returns an answer where every requirement is followed by a [Source: NBC 2020, Page 845, Section 9.10.9.18.(1)] tag that links directly to the page in the source PDF
If the source PDF does not contain a sentence that says X, CodesWise cannot tell you X. That constraint is the entire product.
Side-by-side
| ChatGPT | CodesWise | |
|---|---|---|
| Source of answers | Web-scraped training data, no NBC/OBC/etc. text | Official published code PDFs, ingested + indexed |
| Hallucination risk | High — invents plausible-sounding clauses | Constrained — only cites what's in the retrieved passages |
| Page-level citations | No — cites article numbers it cannot verify | Yes — clickable PDF page-locate on every clause |
| Provincial overlays | None | Ontario Building Code 2024, BCBC 2024, BC Fire Code, Vancouver Building By-law 2025, NBC Alberta, NFC Alberta, Code de construction du Québec |
| Edition currency | Knowledge cutoff, no awareness of 2025 NBC release | NBC 2025 indexed within weeks of publication |
| Audit trail | None — answers regenerate, no record | Every answer keeps its citation chain. Export to PDF for project files. |
| Built for | General use, summarization, brainstorming | Architects, engineers, contractors, building officials, code consultants |
A real example
Question: "What fire-resistance rating is required between an attached storage garage and the dwelling unit it serves under the National Building Code?"
Sounds right. But the actual clause is 9.10.9.18, not 9.10.5. The required rating depends on whether the garage is sprinklered and how many vehicles it serves. ChatGPT cannot tell you which because it never read the clause.
Click any citation pill in the live app and the source PDF opens at that exact page, with the matching text highlighted.
When ChatGPT is fine
For a quick sanity check on terminology, summarising a long meeting transcript, drafting an email, or any task where you can verify the output yourself, ChatGPT works. It is a useful generalist.
For anything that ends up on a stamped drawing, a permit submission, an inspection report, or a compliance memo, you need a tool that grounds its claims in the actual published code. That is what CodesWise is for.