The problem
In a regulated industry, answering a technical question means finding the current, authoritative document, not one of its five ancestors. Searches ran 6 to 8 hours because version truth lived in people's heads, and the highest-value documents extracted worst: scanned pages, spec tables that flattened into noise.
What we built
A narrow-scope retrieval system over a curated corpus: one workflow, one team, one governed document set. Most of the work was cleaning and structuring documents (OCR where needed, version marking, ownership assigned) before any model touched them. That's deliberate: a corpus of documents someone has vouched for beats fifty thousand nobody has looked at.
The result
The same search now takes about 15 minutes, with answers grounded in the current revision. Because the scope was narrow and the result held up, the project earned the credibility to expand, the opposite of the boil-the-ocean assistants that die in the demo phase.