One boundary
Prompt content does not go to a model provider’s API.
Deployment
A local deployment, from the first conversation to a server on your network.
We assemble the assistant, document search and integrations into one boundary that stands at the client. First we agree which data may enter a prompt at all.
B2B
Scoped for a company or an institution, not for a private individual.
No price list
Cost follows the model, the hardware and the integrations, so it is not posted here.
What “stays in the company” means
The model, the excerpt index and the answer service run on a machine the organisation names: at head office or in a server room where you control access.
This is not a shortcut to a “private instance” hosted by a cloud model vendor. A query does not leave the boundary as a call to someone else’s API.
The network may still be needed for system patches or for a backup location you choose. We describe that traffic in the design so it does not carry conversation content.
What the boundary is made of
A typical layout we adapt, rather than a closed box sold as-is:
- An open-source model service on your server.
- A document index, when answers should follow the company’s files. More: document search.
- Accounts and network: who may ask, who administers, what the access log looks like.
- Connections to tools you already use — integrations.
- Training and maintenance described on training and maintenance.
What we do not declare up front
We do not quote tokens per second or a “quality” threshold. Both depend on the weights, the hardware and the quality of the materials.
The assistant does not take legal, HR, medical or financial decisions. A person stays responsible for anything that goes on to a customer or into a file.