Train your agent on your content

Add websites, pages, files, Q&A pairs, and Notion content, then keep them current.

Sources give the agent the facts it can retrieve. Start with the smallest authoritative set. Test it. Then expand.

Add a website

Open an agent. Select Sources.
Choose Crawl website. Enter the site URL. Start the crawl.
Wait for processing to finish. Then test a known question in Playground.

When you need one URL instead of a whole site, use a page crawl. Website and page ingestion need publicly reachable URLs.

Upload a file

Choose Upload file. Select the document. Then wait for processing. The source list shows its status. A file is not available for retrieval until processing succeeds.

Add an exact Q&A

Choose Q&A. Enter the customer question and the answer that you want the agent to use. Then save. Q&A pairs work well for short policies or corrections that do not need a long page.

First, connect Notion under workspace Integrations. Return to the agent's Sources page. Choose the Notion pages to sync. Add them. If the connection needs attention, reconnect it before you retrain the source.

Keep knowledge current

After the source content changes, use Retrain.
Configure Auto retrain for regular website updates.
Remove obsolete sources so that stale facts do not compete with current ones.
Use Retrieval settings to tune how the agent searches its knowledge. Then compare results in Playground.

When you delete a source, retrieval no longer uses it. After a source change, always rerun your highest-value test questions.

Manage sources from an MCP client

The MCP server gives a connected AI client the same source operations as the dashboard:

List sources.
Read a source.
Add a URL or text source.
Rename a source.
Delete a source.
Restore a source.

File uploads and Notion sync stay dashboard-only.

Two behaviors differ from the dashboard, where you can watch a progress bar:

URL sources index in the background. create_source returns as soon as the crawl is queued, with status: "indexing". Poll get_source until the status reads ready. A large crawl takes longer than a single request can stay open.
Deleted sources stay findable for 30 days. list_sources with status: "deleted" returns soft-deleted sources with a restoreBefore deadline. You do not need the id from the delete response to restore a source.