An AI assistant over your internal documentation and knowledge

The symptom is familiar. Someone needs a technical fact, a procedure or the history of a problem, and the answer exists but is buried. They ask a colleague, the colleague asks the veteran and the veteran knows it by heart because the document that holds it is two hundred pages long. Each of these queries steals minutes from two or three people at once, always the same ones.

The real cost is not only the minutes. It is the dependence on a few people, which turns every absence and every retirement into an operational risk. It is the decision taken without the fact, because finding it cost more than the gamble. And it is the new knowledge that stops being written down, because nobody writes for an archive where nothing can be found.

The difference from a search box is who does the final work. A search box returns ten documents where the answer might be, and the person still has to read them. This assistant answers the question, in the language of whoever asked, and attaches the reference of where the answer came from, so verifying costs one click instead of one afternoon.

Underneath sits production-grade retrieval over your documentation, what the sector calls RAG, a search that feeds relevant fragments to the model so it answers from your sources rather than from its memory. The word matters less than the bar. A demo RAG finds the easy things. A production one has to find the hard things, dense documents, house vocabulary and badly phrased questions at three in the afternoon.

Some questions no document answers well, because the answer changes every hour. The state of a machine, the recent history of a fault, a production figure. For those, the assistant does not quote a two-year-old document. It queries the internal system that holds the fact and answers with what is true right now.

That mix is what makes the assistant useful in real work, documentation for the knowledge that gets written once, systems for the knowledge that changes daily. And when the query calls for it, the assistant guides step by step, a diagnosis, a procedure, or executes the action directly on the systems you approve.

Past a certain size, a single agent that does everything becomes one huge piece nobody can maintain. Our architecture for these cases is different. An orchestrator understands the intent of each question and routes it to the specialized agent for that domain, one per area of knowledge. In the industrial case behind this page, that orchestrator coordinates half a dozen specialized agents and the whole adds up to some 86 connected pieces.

The person asking sees none of this. They write their question and get their answer. The architecture matters for what it allows, adding a new domain without touching the others, measuring each agent separately and keeping a failure in one from dragging down the rest.

Trust is not earned by answering well alone. It is earned by answering well and refusing well. There are two different kinds of "no". The "this is outside the product", answered instantly with catalogue text, no search spent. And the "this should be documented and is not", said only after genuinely searching, and leaving a trace, because it is a task for whoever owns the documentation.

Confusing them is expensive in both directions, and separating them takes more engineering than it seems. We tell it in full in its own article.

One design detail that shows how we think. When the assistant offers several options, the person’s short reply, an ordinal, a bare term, a "the second one", is resolved without calling the model, through a simple rule that recognizes those terms and those numbers and that expires after fifteen minutes. It looks like a minor saving and is three things at once, an instant response, one less call to pay for and one less message exposed to malicious instructions hidden in text.

The general rule behind the detail is that the model gets used where it earns its place, understanding open language, and avoided where an ordinary program does the same job faster, cheaper and without surprises.

An assistant that reads all of the company’s documentation raises an obvious question, who may ask it what. Our answer is that the assistant inherits the permissions that already exist. Retrieval respects the role of whoever asks, so nobody receives in an answer a fragment they could not open by hand. And when the assistant queries an internal system, it travels with the person’s identity, not with an all-seeing robot account.

Every decision is recorded as well, what was understood, where it was routed, what was answered. That record is never consulted during execution. It exists so that someone with an uncomfortable question can answer it later with facts.

Everything above runs today at a large industrial company. Operators and supervisors query production and fault data, follow guided machine diagnoses and resolve technical questions, in natural language, on the company’s own infrastructure.

We can show the numbers because we measured them ourselves. Routing calibration, the piece that decides which agent takes each question, was measured on 118 real cases and went from 72 to 91 percent accuracy. We also tried a cheaper model for that piece and the test battery rejected it, ten points of accuracy lost. And we accepted a ceiling on purpose, around 92 percent, because chasing one hundred percent ends up tuning the tests to pass rather than improving the system.

That last decision tends to earn the most trust in a technical committee, because it is the one a smoke seller would never tell.

The assistant runs on your infrastructure, your cloud or your own servers, with your organization’s accounts for access and your data wherever you decide. The repository is in your name from day one. And quality is not declared, it is measured continuously, with a battery of real cases that blocks any change that degrades the system and a weekly watch on what is live.

A system of this size is quoted per project, because the cost depends on your sources, your systems and how much validation you demand. For how the rules of European data protection shape all of this, the full story is on its own page.

For the technical review, this is what sits underneath, one line per piece.