When do articles update for AI agents?

Usually within about 10 minutes of publishing or editing a published article. The embedding pipeline runs automatically — you don’t have to do anything. Translated versions take longer; see below.

What happens under the hood

When you set an article’s status to published (or save changes to one that’s already published), three things happen:

1. The saved article content is available immediately server-side. Anonymous public help center pages may take the short CDN TTL to refresh, so the public page can lag briefly.
2. The embedding pipeline marks the article for a vector index update.
3. A background embedding job runs every 10 minutes and updates the index for every article waiting — sometimes longer for very long articles or when many articles are waiting.

Until step 3 finishes, your AI agents are retrieving from the previous version of the article (or from no version at all, if it’s a new article). After step 3, the new content is live in retrieval.

How to tell if it’s done

The article editor doesn’t show a live embedding status. The fastest practical test:

1. Wait about 10 minutes after publishing.
2. Open Sidekick on a real conversation, or trigger a Capability that should retrieve from the article.
3. Ask a question the new article answers.

If the answer reflects your changes, the embedding has caught up.

What slows it down

• Article length. Once the embedding job reaches an article, a short one embeds almost instantly. A 3,000-word reference takes longer.
• Multi-language. Non-default languages wait for the automatic translation sync, which runs every 15 minutes for every active language. As soon as a translation finishes, that translated version embeds on its own. Allow up to about 15 minutes for non-default languages, more for long articles or a large knowledge base.
• Bulk publish. The embedding job works through waiting articles one at a time, up to 250 per workspace per run. A large batch can take more than one 10-minute run. Allow proportionally more time.

What doesn’t trigger an update

• Editing a draft. AI retrieval only returns published articles. Save freely; nothing reaches AI agents until you publish.
• Editing an article that’s needs-review or archived. Same as draft — no AI retrieval.
• Renaming a category or moving an article between categories. The article body didn’t change, so the embedding stays valid.
• Adding or removing tags / role assignments. Browsing-visibility metadata; doesn’t change retrieval.

Forcing a rebuild

The KB has an admin endpoint to force a full embedding rebuild across all articles — useful after a bulk import or if you’ve noticed the index has drifted. It’s an API-only operation today; there’s no button in the editor. If you think you need it, contact support before running it: full rebuilds take time and rate-limit the AI usage during the rebuild window.

Common gotchas

• You changed an article and Sidekick still gives the old answer. Wait about 10 minutes and try again. If it persists, check that the article actually saved (the editor saves automatically but a stale tab can show outdated state).
• You translated an article and the foreign-language version still says the old text. Translated versions wait for the automatic translation sync, which runs every 15 minutes. After the translation completes, the translated version embeds and becomes available for retrieval. Allow up to about 15 minutes for short translations, more for long ones.
• You unpublished an article and an AI agent still references it. Unpublishing (status → draft / needs-review / archived) removes the article from retrieval within a minute. If an in-flight conversation already retrieved the article before the change, the agent finishes with the old reference. New conversations don’t.

See also

• Knowledge Base search and AI retrieval
• The article lifecycle
• Article statuses