Training data: what the agent knows
An agent answers from what it has been trained on. Training data is the inventory of that: every document the agent's index holds, what state each one is in, and three ways to add more.
The page is at Aura AI → Training data. It is not in the section's navigation column — the way there from inside the product is the Teach the agent link that sits beside every unanswered question on the Performance page, which is the moment you usually want it. It is a real page and you can bookmark it.
Adding, re-indexing and deleting need admin or owner. Anyone can read the list.
The list
The heading is Indexed documents, with a census line underneath: how many documents, how many indexed, how many failed, and when the list was read.
Each row is one document: a status pill, its title (or its address, or "Untitled document"), its source type, a link, when it was last updated, and its data types. The status pill is one of Indexed, Indexing, Failed or Unknown. Unknown means the index did not report a status, which is not the same as a failure. A failed row shows the index's own reason for failing, or, when it gave none, says that Aura AI could not index it.
Above the list: a search box over titles and addresses, filters by source type and data type, a sort (Newest first, Oldest first, Recently updated, Title A–Z) and a page size of 25, 50 or 100.
Occasionally the list warns that it is not showing everything in the index. It says so rather than presenting a partial list as complete. Narrow with a filter or a search when that appears.
There are three different empty states and the difference between them matters: "This agent has not been trained yet" is a new agent, "No documents match this search" is your filters, and "That document has been deleted" is the confirmation after you removed one.
Adding sources
Add sources opens a dialog with three tabs.
Upload files. PDF, DOC, DOCX, TXT, CSV, JSON, XLSX or Markdown, up to 10 MB each. Drop them in or choose them; indexing starts as soon as a file is accepted. Each file reports its own progress — waiting, uploading, queued for indexing, or not uploaded — so one rejected file in a batch of twenty is visible rather than silently missing later.
Web pages. Paste addresses, one per line. Auralis lists anything it cannot read as an address so you can fix it, and tells you how many are ready before you send them.
Crawl a site. Give a site address and Aura AI looks for the pages it has and lists them. Nothing is indexed until you choose from that list — the crawl is discovery only. You then tick the pages you want, fifty to a page, and submit those.
If the crawl hits a site behind a sign-in, Auralis says so and suggests the workable alternative: crawling cannot get past a login, but if you can reach those pages yourself you can paste their addresses under Web pages.
Re-indexing and deleting
Two row actions, both for admins, both confirmed.
Re-index from source fetches the document from its source again and re-indexes what comes back. It removes the current copy first, so what returns is a new entry in the list rather than an update to the old row. Do not be alarmed when the timestamp resets; that is the mechanism.
Not everything can be re-indexed. A document without a fetchable source, or one whose kind does not support it, is refused with a sentence saying which of the two it is, rather than failing silently.
Delete removes the document from the agent's training data and the agent stops answering from it. The deletion happens inside Auralis itself, not only in this list, and it cannot be undone from Studio. To bring the content back you add the source again.
A note on help-centre articles
Articles from your own help centre are not added here. They go to the agent from Knowledge → Content, where each article carries its own AI state and the selection is made against the articles themselves. Two inventories of one index is one too many. See Sending articles to the AI knowledge base.