The System Learns While You Work
The system drafts what it can, you correct it while you do your job, and your correction is what the next person gets.
Knowledge management has always failed at the same step. It asks you to stop working and write down what you know, and nobody has a spare hour at the end of a hard day. V.E.T.S. removes that step instead of asking harder. This page describes the loop that replaces it, what runs to close the loop, and where the seams are.
The knowledge flywheel
Four things happen, in order, and then they happen again. Each pass starts from better material than the one before it.
1
The AI writes the first draft. You do not start at a blank field. The system generates a first pass and puts it in front of you. Its own knowledge base was seeded this way: 270 of the descriptions in it are generated accounts of stored procedures, and another 37 describe tables. A wrong draft is far easier to fix than an empty page is to fill.
↓
2
You correct it while you work. The correction happens where the work happens. Nearly every clinical record type has a mobile twin built for a phone held in a barn aisle — 27 of the 33 farrier procedures are astp_Farrier_Mobile_ procedures. You are not writing documentation. You are fixing a wrong line in a record you had open anyway.
↓
3
The system takes the correction back in. A scheduled job carries changed content into the search index every hour. A second job runs once at 02:00 and does the slower work: summaries, keywords, relationships between documents, and a recalibration of how much each one is trusted.
↓
4
The next person gets the corrected version. Not the draft. The person who benefits is usually not the person who did the fixing, and usually does not know a fix happened. That is the point.
↻ and the next draft starts from better material
Step two is the load-bearing one. Every other knowledge system in this field also wants your expertise written down; what makes this loop turn is that writing it down is not a separate task you could skip. If correction were a chore, the flywheel would stop on the first busy week.
Why asking people to document more has never worked →
When your correction becomes searchable
Fixing a record changes the record immediately. Making that fix reachable by a question someone asks later is a second thing, and it does not happen in the same instant.
The job that carries new and changed content into the vector index is ajob_AI_GCPVectors, and it runs every hour, on the hour, all day. Separately, ajob_AI_Intelligence_Pipeline_Daily runs once at 02:00 and does the heavier curation that step three describes.
So there is a seam, and it is worth naming. Between your edit and the next hourly run, the corrected text is in the record but is not yet in what the AI retrieves. Open the record and you see your change. Ask a question that should surface it and you may get the old answer, for up to an hour. This is a timing window rather than a queue — the only thing waiting at any moment is whatever changed since the last run.
How a correction becomes something the AI can retrieve →
Why the loop needs a person in it
An AI first draft is a draft. The 270 generated procedure descriptions and 37 generated table descriptions in this system’s knowledge base are useful because people who know those objects have been correcting them, not because a model produced them.
This matters more than it sounds. A system that learns only from its own output drifts, confidently, in whatever direction it started. The expert’s correction is the only genuinely new information entering the loop — everything else is the system rearranging what it already had. Remove the person and you do not get a slower flywheel. You get a fast one spinning on nothing.
It is also where the value changes form: a recorded observation is not the same thing as a judgment about what to do next, and the platform is deliberately built to hold both.
Where an observation turns into a judgment →
What is running, and what is not in the index
This is not a prototype. The platform holds 47,062 animal records across its tenants, and the loop above runs against a working system rather than a demonstration one. Fourteen assistants cover 78 different page contexts, so the help you get is scoped to the screen you are already on.
The searchable index that the AI reads from holds 2,333 documents, and it is worth being precise about what those are, because most readers assume the worst:
The index holds documentation about the system. Help text, knowledge-base articles, and generated descriptions of database objects. No animal records. No client details. No patient history. Your records are never embedded and are not in the index.
That is a real boundary, not a wording trick. An assistant working alongside you on a live screen still reaches your actual records through the application’s ordinary paths, subject to the ordinary permission checks. That is a different mechanism from the index described above, and the two are kept apart deliberately.
Who can see which records, and how that is enforced →
Where to go next
The flywheel is the shape of the thing. If you want to watch it turn once, in order, the next page follows a single visit from the first record entered to the answer someone else gets from it later.
Follow one visit end to end →