How it works: search first, write afterwards
When the system is set up, the documents are split into pieces and stored in an index that allows searching by meaning, not only by words. When someone asks, the system finds the relevant fragments, passes them to the model along with the question, and the model writes the answer citing where it comes from.
This is called retrieval-augmented generation. The advantage is twofold: the model does not need training, and when a document changes it is enough to index it again.
Which documents work
PDFs, text documents, spreadsheets, presentations and email all work. Scanned ones first need text recognition. What goes in badly comes out badly: a crooked scan or a table split across two pages will give worse answers.
Permissions: everyone sees their own
A business assistant has to respect folder permissions. Someone who cannot open the payroll folder must not receive an answer drawn from it either. Ask for it from the start, because it is easier to build it right than to fix it later.
What hardware it needs
Three pieces are at work: the language model, which is the one that needs graphics memory; a small model that turns text into vectors; and the index, which lives on a fast drive. The number of documents hardly changes the hardware. What changes it is how many people ask at once.
| Case | Hardware | Price |
|---|---|---|
| Small team, large model and silence | Compact AI 64, unified memory | €2,940 |
| Up to 15 people | Ryzen 9 9900X · 64 GB · RTX 5070 Ti 16 GB | €4,482 |
| Up to 30 people | Ryzen 9 9950X · 64 GB · 2× RTX 5070 Ti 16 GB | €6,155 |
| Up to 60 people | Ryzen 9 9950X · 128 GB · 2× RTX PRO 4000 Blackwell 24 GB | €12,252 |
What you should know about its limits
The model answers with what it finds. If the document is out of date, so is the answer. If the question requires cross-checking twenty documents, it may miss one. And if it finds nothing, a well-built system says so instead of making something up. That is why the citation matters: it lets you check the answer in ten seconds.
Related hardware and guides
Frequently asked questions
Is the AI trained on my documents?
No. The documents are indexed so they can be searched, but the model does not change. If you remove a document from the index, it stops appearing in the answers.
How many documents can it handle?
Tens or hundreds of thousands without trouble: the index takes little space and lives on a fast drive. The size of the hardware is set by the number of people asking at once, not by the number of documents.
Does it respect who can see each folder?
It must, and you have to ask for it. The assistant checks folder permissions and only uses what that person is allowed to open.
Can it be wrong?
Yes. That is why every answer carries the citation of the document it comes from. If a fact matters, you check it in the original, just as you would with a new colleague.
Shall we do the sums for your company?
Tell us how many people will use it and with which documents. We will reply with the hardware, the installation and a fixed price within 12 working hours.