Connect the desk

The wizard does this for you. This page is what it does, for when you would rather do it yourself.

Two tables in nils-desk.toml. The desk proxies both under its own origin, with the person’s own token, so a browser never holds a key and neither Kvasir nor the assistant faces a network.

[kvasir]
url = "http://127.0.0.1:7100"

[assistant]
url = "http://127.0.0.1:7300"

Restart the desk. It shows the assistant to a person who may use it, as their groups and grants say, and to nobody else, and it shows nothing at all if the assistant did not answer. With [kvasir] set, Settings has the Language models page, where a person who uses the assistant and may see that page adds their own ChatGPT subscription.

What to expect

The first request after a model starts is slow, because the model is warming; the desk says so rather than looking broken.

Everything the assistant proposes arrives as a change with a diff. Nothing is applied until a person accepts it, and what a person rejects is remembered as a correction.

What leaves the machine

With a local model, nothing: the prompts, the schema and the counts stay inside the deployment.

With a commercial provider or a ChatGPT subscription, the prompt leaves. Kvasir marks such a backend remote and refuses the purposes nobody has opened to it, so that decision is made once, on the Language models page, by a person with work there, rather than in each station. A purpose that reads rows of the registry opens only with a written reason, and identifiers never leave.

Without the assistant

Everything else keeps working. The desk shows what is there; a part that is absent is simply absent, and no page waits for it.