Canadian Codes
A retrieval search product over Canadian building, fire and electrical codes, federal to municipal, that answers in plain english and opens the exact provision it used. Because when an answer MUST be correct, you can't risk a predictive answer.
The codes are public and free. Finding the clause that actually applies to your project is the hard part, and getting it wrong means building something you have to tear out. Ask around a job site or an LLM and you'll get a confident answer that's either a decade out of date or from the wrong jurisdiction.
Finding the right passage means slicing thousands of pages into searchable pieces. Slice by length and you cut rules in half: a clause depends on the heading above it, much of the content lives in tables, and the same clause number means different things in different provinces.
So pieces follow the document's own article boundaries, tables kept whole. That took the collection from 116,000 pieces to 36,000 and exposed 16,800 duplicates that length-based cuts had been hiding. Search runs keyword and vector together, narrowed to one jurisdiction before it runs. Postgres won't index vectors past 2,000 dimensions and mine are 3,072, so they are stored at half precision: under the ceiling, no meaningful loss of accuracy, and an index that exists at all.
The obvious objection is that a model with a million-token context could just read the whole code book. It can, and it still cannot tell you which clause it used. Every answer here opens the provision behind it, which is the difference between an answer you trust and one you can check. Running for paying subscribers, accepted into UofT ICUBE and Microsoft for Startups.