Add knowledge sources
Give the assistant selected public websites and linked documents to answer from.
On this page
What can be a knowledge source?
Administrators add sources in Knowledge. The source must be publicly accessible without signing in. Runs Agents can render webpages and extract supported public documents found within the source scope. There is no private-file upload or authenticated document fetch workflow.
A source URL establishes a hostname and path boundary. A source at https://example.com/docs covers that path and its descendants, rather than unrelated sections or other subdomains. The bare hostname and its www counterpart are treated as aliases. Overlapping configured source scopes are rejected.
Adding a knowledge source does not install the chatbot on that website. Website display is controlled separately in Installation.
Add and inspect a source
- Open Knowledge and choose Add source.
- Enter the public website URL.
- Choose whether to receive first-setup email updates.
- Let discovery continue in the background, then open the source.
- Review the page hierarchy, selection, and statuses.
Sources with at most 1,000 discovered URLs start activation automatically. Larger sources pause for selection review and Activate source. The threshold does not cap manual activation at 1,000 pages. Discovery has a supported ceiling of 100,000 URLs, counting webpages and documents together; a source that reaches it can have an incomplete inventory.
Supported public documents
| Format | Support and practical requirements |
|---|---|
| Text extraction; do not rely on image-only or scanned content being read. | |
| DOCX, PPTX, XLSX | Extractable text and supported table content. Images and embedded attachments do not become knowledge. |
| TXT | UTF-8 text. |
| CSV | Readable tabular data within extraction limits. |
Documents must remain within the public source's hostname and path scope, including redirects. A document download is limited to 5 MiB. Additional extraction limits can make a complex or highly compressed file unavailable even below that download size. A recognized extension alone does not establish successful extraction. Check the page's actual status.
Make public information usable
Publish facts as ordinary text with descriptive headings. Keep qualifications beside the claims they qualify. Link important pages and documents from within the source scope. A crawler cannot use an unlinked private file merely because its name is mentioned in instructions.
If robots.txt blocks CloudflareBrowserRenderingCrawler, the website owner must allow it before discovery or refresh can proceed. Runs Agents does not edit the website's robots.txt. Password prompts, bot challenges, inaccessible links, empty output, and unsupported documents can also leave pages unavailable.
Remove a source
Deleting a source removes its indexed website content. Conversation records are managed separately. See Retention and deletion.
Continue with Select pages and update knowledge.