AI that answers from your documents or does not answer at all.

The problem it solves

Every company holds knowledge that is in no system: procedures in PDFs, what was agreed with a customer a year ago sitting in a mailbox, a price list in a spreadsheet, an exception to the rule in one long-serving employee's head. A new hire finds out by asking whoever is available. When that person is away, the company slows down.

A language model connected to the internet does not solve this, because it has never seen your documents. A model given your documents without rigour solves it in appearance - and then invents, confidently, in the few percent of cases nobody checks.

How we build it

Retrieve first, answer second

The question goes to a search across your documents, and only the passages found go to the model. It does not answer from memory, it summarises what it was handed. If the search finds nothing there is nothing to summarise, and the system says it does not know.

A citation, not a paraphrase

Every answer carries a link to the document and page. You click and see the original. That is the whole difference between a system you can defend in an audit and a demo.

Extraction into fields, not prose

Where specific values have to come out of a message - route, amount, deadline, reference - the system returns typed fields rather than a paragraph to read. A missing field is marked missing, and that is information rather than failure.

A trail behind every decision

We record what was asked, which passages were used, which model answered and how long it took. Without that you cannot improve quality, because you cannot see what is breaking.

What it will not do

It will not fix disorder in your documents. If three versions of the same procedure are in circulation and nobody knows which is current, the system finds all three. That is often the most valuable outcome of the first week - and often the first piece of bad news.