A technician calls in a fault. Whoever picks up starts searching the shared drive for anything that sounds like it — a service note, an old ticket, a page from a manual. Nine results come back. The one at the top is long, detailed, and uses almost the exact words the technician just said. It is also about a part that was replaced a year and a half ago.
The note that actually applies is three results down. It is shorter, plainer, and correct. Nobody picked the wrong one on purpose — the search box does not know which document is still true, only which one uses more of the same words.
What should annoy you is not that an old note exists. Old notes always will. It is that nothing on the page said so, and the only way to find out was to read both and guess.
Why we build it this way
The value in a plant's own history was never the model reading it. Anyone can rent a model. The value is the drawings, the service records, the tickets and the notes that only exist inside that one company, and almost nobody has made that pile usable.
Most attempts treat this as a search problem: put everything in one box, index the words, return the closest match. That works in a demo, where the only document in the box is the right one. It falls over the moment two documents disagree, because a keyword index cannot tell you which one supersedes the other. It can only tell you which one used more of your words.
The problem is not text. It is relationships. A drawing has a revision, and a revision has a successor. A ticket is about a specific part, not a part number that later changed hands. A service note was written by someone, on a date, about a machine that has since been modified. None of that lives in the words on the page. It has to be built in on purpose.
A search box tells you what matches your words. It cannot tell you what is still true.
How it works, roughly
We build a graph over what a plant already has — parts, assets, people and documents — and the relationships between them, so a revision knows its predecessor and a ticket knows which part it was actually about. It reads whatever exists, in whatever state it is in:
- Drawings, and which revision superseded which
- SOPs and service records
- Tickets, and the fixes that were eventually attached to them
An answer comes back with the document it was drawn from, named and dated, not just a page number. Anyone reading it can open that exact source and check it before they act on it.
When a person disagrees with the answer, nothing changes on its own. They open the underlying document, and if it turns out to be wrong or superseded, that correction goes back into the graph — a person's judgment, not the system's, decides what the record says next time.
A year in, the question "which document is this even from" stops being a research project. The graph stays inside the plant's own building, and whether an answer is good enough to act on is still a person's call.
Read about AI Transformation →