Add files and public websites that your agent can retrieve as cited reference material.
A knowledge base gives an agent reference material it can retrieve while answering. It is useful for stable policies, product documentation, procedures, and help-center content. Use client tools for live or per-user data such as an order status or account balance.
Knowledge is configured per agent. Enabled, indexed content is retrieved with hybrid semantic and keyword search, and answer citations retain the source title, URL when available, and matched text.
Use the actions at the top of the knowledge library:
The library updates as work progresses. Search, sort, or filter it by type and status, then open a source in the detail drawer. File sources show their indexed Content; Pages, Settings, and Activity show discovered pages, source settings, and sync history.
Files may be up to 10 MB. They are normalized to Markdown before indexing; sources whose normalized content exceeds 4 MB cannot be indexed. Check the source status and extracted content after upload rather than assuming a particular document layout will produce useful text.
Website sources crawl public pages only. Pages that require authentication are not available. The service respects crawl restrictions and does not fetch content from your browser.
Choose a parent folder when creating a source, or use its action menu to rename or move it later. Select multiple entries to move them together. Folder breadcrumbs let you return to a parent folder or the library root.
Folders only organize the library; moving a source does not change its indexing or whether the agent can use it. Folders belong to the same agent as their sources. A folder must be empty before it can be deleted. Deleting a source requires confirmation and removes its content from the knowledge base.
Website settings apply to a source, not just its root page:
* does not cross /; ** does. Excludes win.Use Test matching before a large sync to see whether representative URLs will be fetched and indexed. A source retains the URL and title for each child page, so you can exclude or retry a page individually.
Preview page crawls and shows one page with the current settings without indexing it. Sync starts a fresh source run. Website schedules are manual, every 6 hours, 12 hours, daily, or weekly.
These controls are intentionally separate:
During a replacement sync, the previous ready revision remains available if a page or indexing step fails. Review a failed or partial run, then retry the affected page or sync the source again.
The workspace owner receives an email when a newly added source, manual sync, or retry finishes. The summary covers that source—not the entire knowledge base—and includes indexed, failed, and skipped page counts, with a link to review the source.
Successful automatic refreshes stay quiet. Partial or failed automatic refreshes send a summary so the owner can review the run. Previews and cancelled runs do not send indexing emails. Notifications require a verified owner email address and apply to runs started after this feature was enabled.
An indexing failure does not remove the previous ready revision. Disabled sources remain unavailable to the agent even after indexing succeeds.
Source runs move through Queued, Crawling, and Indexing, then finish as Completed, Partial, Failed, or Cancelled. A source is Ready when it has usable indexed content; Partial means some pages failed while other ready content remains. Deleting hides the source immediately while removal finishes.
Child pages show Queued, Processing, Indexed, Skipped, Failed, or Cancelled. A skipped page was intentionally outside the selected scope or rules. Open its details to see the reason, extracted Markdown, URL, and title.
Turn on the Search knowledge toolkit under Tools. The agent automatically retrieves relevant passages before answering, avoiding an extra model round trip to request a search. Only relevant excerpts are returned, not entire sources. If the initial lookup is empty, fails, or times out, the agent can still search during its response.
Sources in an answer cite their child-page URL and title for websites, or the uploaded file name for files. Keep source names and page titles descriptive, and keep headings in the original content so citations are easier to inspect.
Select Playground beside the Knowledge Base heading to ask your agent questions without leaving the library. The full-height chat uses the agent's saved instructions, model, and tools, with source citations enabled. It does not include the appearance preview's demo tools or demo page context.
The playground uses enabled, indexed sources across the entire agent, regardless of the folder or filters open in the library. Enable Search knowledge under Tools and index at least one enabled source to begin. Use New chat to test a question without earlier conversation context. Preview chats are excluded from monthly active users; model and tool usage still applies.
For new integrations, use the source-based Knowledge API. It exposes the same sources, folders, child pages, runs, previews, rule tests, retry, and cancellation controls as the dashboard. Earlier document endpoints remain available during the migration.