The model (AI)#

Hyvop works without a language model. AI is off by default, and until an administrator sets one up, nothing is sent to any model host. A notice above every screen says so: "AI features are off until a model provider is configured".

What needs a model, and what does not#

Ownership suggestions that come from the cloud's own facts and from a deployment manifest need no model either (ownership.md).

Which model answers#

There are two places a model can come from.

The install's model. The administrator who installed Hyvop names a provider and a key in the chart (core.env.AI_PROVIDER). See models.md. It is the default for every organisation.

The organisation's own model. Settings > LLM Configuration (/settings/llm, administrators). For each tier, good and small:

• the provider: OpenAI, Azure OpenAI or an OpenAI-compatible endpoint;
• the address, where the provider needs one;
• the model, or for Azure the deployment;
• the key. It is stored encrypted, never shown again.

Then Test. The test sends one short question. A saved model is not used until it has passed; a failed test keeps it out of use and shows the reason. Changing the provider, the address, the model or the key clears the test. The test is counted like any call, and refused when the month's budget is used up. Remove takes the model away.

Once an organisation's model has passed, every AI call of that organisation goes to it. Until then, and for an organisation that brought none, the install's model answers. The screen shows the model in use now, and whether it is the organisation's, the install's default, or none (AI off).

An address inside your network, such as an Ollama or vLLM server, must be allowed by the administrator (core.env.OUTBOUND_ALLOWED_HOSTS); a public HTTPS address is accepted as it is.

The assistant#

The assistant button appears only where a model is set up. It streams its answer, names the model that answers, and says that the question goes to it. It remembers the conversation of the person signed in, per channel, and no one else's. It reads and explains with the tools it has. It cannot delete or change a resource, and it is told never to make up a figure.

Budgets, timeouts and limits#

Every call to a model goes through one gate that checks:

• The token budget. There is a budget for the month for the organisation and a budget for one run (a knowledge-graph build, one investigation). A call that would go over is refused with a sentence that gives the numbers. A run that reaches its budget stops and says why. Zero means no limit.
• The deadline. A call that takes longer than 120 seconds as shipped is cut.
• A limit on the answer. A call with none of its own is capped, 4096 tokens as shipped.

The install's defaults are set in the chart (models.md, "Limits on use"). Budgets an administrator sets in the screen belong to the organisation and win over them.

Usage#

Settings > AI Usage (/settings, administrators). Choose 7, 30 or 90 days. It shows the number of calls, tokens in and out, how many calls were stopped by the budget and how many timed out, and a table by feature. The estimated cost is shown only when the administrator has given prices. The month's budget form is on the same tab, with the month's use against the limit.

What is sent#

Hyvop sends no rows of your data and no names of people, unless an administrator chooses to. ownership.md has the rules and the ownership steps. The others:

Everything above goes only to the model the administrator named, at the address and with the key they gave.