AI agents
AI assistant over your documents
Your teams query your procedures, contracts and files in plain language, with sources cited, without your data leaving your server.
Who it is for
- A firm or SME whose knowledge sleeps in folders nobody can find.
- A team answering the same internal questions ten times a day.
- A company that refuses to send its contracts to a US provider.
What you receive
- Document ingestion and chunking, vector database.
- Chat interface (web, Telegram or Slack) with source citations.
- Access management per team.
- Fully self-hosted option on a local model.
- Test question set and answer checks before opening.
- Documentation and training.
How it goes
01
Scoping
Which documents, for whom, with which real questions: we start from your actual requests.
02
Build
Ingestion, index, interface, on a sample of documents.
03
Checks
Question battery, verified answers, confidence thresholds tuned: the assistant says “I do not know” rather than inventing.
04
Opening
Roll-out to the team, access, follow-up of unanswered questions.
What we have already done
A self-hosted model runs on our own server for tasks where data must not leave; the assistant relies on the same infrastructure as our agents in production.
What it costs to run
Self-hosted, the marginal cost of a question is close to zero. Through an API, every question is measured, capped, and the model comes from our routing table: sensitive data never goes to a provider outside the agreed frame.
Frequently asked questions
Can the assistant make up an answer?
It is tuned to cite its sources and to answer “I do not know” below a confidence threshold. We paid once for a model that invented; it became a rule.
Do my documents go to the cloud?
Not if you choose the self-hosted option: everything runs on your server or ours, in Europe.
How many documents?
From a few dozen to a few thousand. Beyond that, we split by team and by use.