Handing articles to the AI
The include flag, the bulk train action, the two preconditions, and the states an article moves through.
Your help centre and your AI agent's knowledge base are two different stores. Publishing an article does not put it in the agent's index, and nothing puts it there automatically. Handing an article over is a deliberate act, and this page is about that act.
There are two controls: an include flag on each article, and a bulk action over rows selected on the Content list.
Two things have to be true first
A connected Aura AI agent. Articles go to this workspace's connected agent,
resolved server-side from that connection on every submission. There is no
destination field anywhere on this path — not on the form, not as a query
parameter — so you cannot point a send somewhere else from here. With no agent
connected, the include toggle is disabled and says so, with a link to
/auralis-ai where an owner or admin connects one. Nothing is sent, and no
request is even issued.
A public address for the help centre. This is the constraint that surprises people, so it is worth being blunt about the mechanism:
The indexer fetches your articles over the public web
It is handed the article's URL, not its text, and then visits that URL like any other reader.
Two consequences follow. A help centre with no public address has no URLs to hand over, and nothing can be sent until you set one — see Branding and domain. And an article that is a draft, hidden or internal cannot be indexed at all: the indexer arrives signed out and gets the same page as content that does not exist. Those are refused before sending rather than being allowed to fail later.
Only a published, public article can go to the AI.
The include flag
The article editor's details rail carries Include in the AI knowledge base.
It is an exclusion flag rather than a send button. Turning it on does not hand the article over; it puts the article in the set that a later, explicit send is allowed to include. Turning it off marks the article as not belonging in the index.
Excluding never deletes anything. If an article is already in the index when you exclude it, it comes back flagged Remove from AI — a fact to act on, not an action that has happened. Removal is a separate, deliberate call.
Training from the Content list
Knowledge → Content shows every article with its AI state. Tick rows and one selection bar appears, stating the number of ticked rows on the left and carrying four verbs: Delete, Move, Remove … from AI and Train … with AI.
The bar states the selection; each button states its own scope. A mixed selection shows Train 12 of 30 with AI, because an article must be published and public to be trained and the other eighteen stay as they are.
There is deliberately no "train everything" button. What is selected is what is sent, and the selection follows you across pages, so you can filter, page and tick without losing it.
Confirming opens a dialog that states both facts together before anything leaves: the destination agent by its full identifier, and the exact count. The extra keypress is the point — this is the one action on the page whose blast radius is another company's search index.
The report
The banner afterwards gives you sent, failed, refused and not-yet-reached separately, never one number:
- Sent — handed over, with the destination named in the title.
- Failed — the service's own words are shown.
- Not sent because they cannot be — each with its reason. They never left.
- Not reached — one send carries a limited number. Press again to continue: articles that have never been sent go first, then the ones sent longest ago, so a second press continues rather than repeating.
It ends with the sentence that matters most: sent is not the same as searchable. The AI service confirms indexing separately, and the AI column updates when it does.
Removing
Remove … from AI is the only action in the product that deletes from the index, and it has its own bar and its own count because "should be trained" and "needs removing" are usually opposite sets. The dialog says, in those words, that it cannot be undone from here: putting an article back means training it again and waiting for the service to re-index it from scratch.
An article can only be removed from Indexed, Out of date or Failed. While the service is still writing an article, a deletion could be undone by the indexing already in progress, so the server refuses one.
The states
The AI column shows one state, and beside it any flags that are true about it.
| State | What it means |
|---|---|
| Not indexed | Never handed to the AI knowledge base. |
| Pending | Handed over. Either the service has not been asked about it yet, or it accepted the article without confirming it is indexed. Accepted is not the same as retrievable. |
| Indexing | The service reports it is still working on it. |
| Indexed | The service confirms it holds the article and can answer from it. |
| Failed | The last attempt did not work. The service's own words are shown with the article. |
| Removing | A deletion was accepted and is not yet confirmed. Accepted is not the same as removed. |
| Removed | The service no longer lists the article, so it cannot answer from it. |
Three flags sit beside the state rather than replacing it, because an article can be in the index, out of date and excluded all at once:
| Flag | What it means |
|---|---|
| Out of date | The version handed over is not the version published now. |
| Remove from AI | In the index and no longer allowed there — unpublished, no longer public, or excluded. Nothing has been deleted. |
| Excluded | Somebody marked it as not belonging in the index. |
Out of date is the one to build a habit around
It appears whenever you edit and republish an article that is already in the index, because editing does not re-send it and nothing else does either. Sending again is always something a person does, and the flag exists so you can find the articles that need it.
The Content list's + Filters disclosure carries an AI knowledge base chip group with exactly these two derived flags: Out of date and Needs removing. Both are computed on every read by comparing the article as it stands now against what was handed over, which is why they are filters rather than editorial states.
A workable routine
Train in one pass after your first publish. Then, once a week or once a sprint, filter Content by Out of date, tick everything, and train again. That is the whole maintenance story, and it takes a minute.