Aller au contenu

The page assistant

Ce contenu n’est pas encore disponible dans votre langue.

Open an Adminium page that builds documents — Email templates, Report builder — and there is an Ask button in the header. It opens an assistant that already knows what that page holds: the templates in it, the format they are written in, your branding, and which tables your role can read.

It drafts. It never saves. Every button that would change something is locked until you turn actions on for that session, needs the same permission the page’s own Save needs, and asks once more before it runs.

Page Where
Email templates, and its editor In the header, before Duplicate
Report builder, and its editor In the header, before Duplicate

The button appears only for a role that holds the assistant permission. It does not disappear when no AI provider is configured — the window opens and tells you what is missing and who can fix it, because a button that vanishes teaches nobody anything.

Always:

  • The page’s own documents — names, and the one you have open.
  • The document format the page accepts, generated from the page’s own validator, so a draft that passes there is a draft the editor can open.
  • Your database schema — tables, columns and their types, for the tables your role can read.

Only when an administrator turns it on:

  • Rows from those tables. Masked, at most 50 per request, and every table a turn touched is listed under Sources read on the result.

The switch is in Settings → AI, off by default, and it is not carried by an exported bundle: importing somebody else’s configuration can never turn it on for you.

Everything it reads goes to the AI provider you configured. If that matters for your data, leave the row switch off — the assistant still drafts from your documents and your schema.

  • It never writes. The model’s last move is a draft. Creating the row is a separate action you take, behind a confirmation.
  • It never sends anything to your customers. Send test email goes to your own address and nowhere else.
  • It never publishes. A saved draft is a draft: an email template arrives disabled, a report arrives with status draft.
  • It never changes existing rows. A draft changes no customer record, and the window says so.

The assistant is gated by a system permission, seeded on a fresh install to Super Admin and Admin — not to Editor or Viewer. Opening it sends your documents, and (with the switch on) rows you can read, to a third-party model; withholding that from a role is a decision an operator has to be able to make. Grant or revoke it on Studio → Roles.

Saving is a second permission — the same one the page’s own Save button needs, which on a stock install only Super Admin holds. So a role that can draft but cannot save is the ordinary case, not an edge:

Your role can look, draft and preview here, but not save.

Everything still works — questions, drafts, previews, and (in an editor) putting a draft straight onto the screen so you can save it yourself. Only the buttons that write are locked, and they say why.

Every action that changes something writes an audit row, so who asked for this draft, and who saved it is answerable months later. The window says so in its footer. Reading is not audited; saving, sending a test and adding a language variation are.

Each conversation is stored as a session with its turns, and swept away on the retention schedule you set for the workspace.

The assistant uses the same AI provider Adminium’s schema enrichment uses — configure one in Settings → AI. There is no bring-your-own copy-paste path here: a conversation is many round trips, and pasting each one is not a workflow anybody would want. LLM assist covers the enrichment paths, where BYO is a first-class option.

What you see What it means What fixes it
No AI provider is configured yet. Nobody has set one up on this instance. An administrator configures one in Settings → AI.
Outbound network features are off on this instance. This install is air-gapped, or outbound calls are switched off. A change to the instance’s configuration — the assistant needs to reach a provider.
You do not have permission to use … Your role does not hold the assistant permission. Whoever manages roles grants it on Studio → Roles.

Settings → AI also decides what the assistant is called. The name shows on the Ask button and throughout the window. It is per instance, so a workspace can call it whatever suits the people using it.

The reasoning behind the shape of all this is on the assistant proposes, a person saves.