How AI support works

Agents that answer from your server’s own knowledge.

Bao's AI support is built on agents. An agent is four things:

a personality (its name, its avatar, the way it speaks, its scope);
knowledge (your documents, your pages, your channels);
a model (its reasoning ability, and its cost);
behaviours (where it steps in, what it is allowed to do, when it hands over).

A server can have several agents. For instance a cheerful greeter in public channels, and a technical expert in tickets, each with its own identity.

What an agent is allowed to say

This is the most important question, and the answer is not "only what is in the knowledge base".

Kind of information	Allowed source
Facts specific to your server: prices, procedures, lead times, rules, accounts	Your knowledge only. Inventing them would be a fault.
General knowledge: how a piece of software works, a technical notion, a piece of advice	The model's general knowledge, freely.

Forbidding everything not in your documents would turn the agent into a stupid gatekeeper. An agent on a PC-optimisation server must be able to explain how to switch off Windows telemetry, even if you never gave it the document that says so.

Restricting an agent's scope ("you only talk about optimisation") is done in its personality, not elsewhere.

Strict mode exists for the rare fields where any statement outside the base is a risk: legal, banking, medical. It is off by default.

What stays out of an agent's reach

An agent can only do what you explicitly allowed. An action that is not enabled is never offered to the model: it does not exist for it. There is therefore no magic phrasing that would trigger it.

Beyond that, whatever the model says, Bao systematically neutralises @everyone, @here and role mentions in its replies, and caps their length.

AI support is included from Bao Royal onwards.