Knowledge bases
A knowledge base is a named body of content the agent searches while it answers: a crawled site, uploaded files, or both. Each one is made of sources, indexed on their own. Add a knowledge base to an agent and it searches the sources that are ready, answers from the passages it finds, and cites them under the answer.
Create one and add sources
Section titled “Create one and add sources”On Knowledge, press New knowledge base, name it, and press Create knowledge base. Then add sources on its page:
- Add URL crawls a public page and the pages under its path, such as a help center’s home. Crawling starts at once.
- Upload file indexes one file as a source of its own. Plain text, code and markup up to 10 MiB; PDF, Office files, HTML, CSV and images up to 4 MiB. Images aren’t read for text.
Each source shows Indexing with its progress, then Ready with its page count, or Error with the reason. The Try a question box on the knowledge base’s page runs the same search the agents run, so you can check what it finds before an agent does.
Crawl options
Section titled “Crawl options”Set them when adding a URL, or later from the source’s menu under Edit crawl options. Saving options doesn’t crawl again; Sync the source afterwards.
| Option | Default | What it does |
|---|---|---|
| Page limit | 500 | The most pages to index. 500 is also the cap. |
| Link depth | Any | How many links to follow from the URL. |
| Discovery | Sitemaps and links | Where the crawl finds pages. |
| Include patterns | Under the URL’s path | One per line. * matches within a path segment, ** across. |
| Exclude patterns | None | One per line. Matching pages are skipped. |
| Render JavaScript | On | Needed for pages built in the browser. Off is faster. |
| Include subdomains | Off | Also crawl the same path on the site’s subdomains. |
A site’s Content Signals can refuse the crawl, which fails the run with that reason.
Every sync is a run, shown newest first on the source’s page with its stages: Crawl, Store and Index for a URL, Upload and Index for a file. A run lists every page it went through with Indexed, Failed and its reason, or Not indexed.
While a run goes, what the source indexed before stays searchable. When it finishes, unchanged pages are kept, changed pages are replaced, and pages the crawl no longer found are removed. Cancel run stops it and deletes nothing.
Give it to an agent
Section titled “Give it to an agent”Open the agent’s editor, press Add a knowledge base under Knowledge, and save. The agent’s prompt then names its knowledge bases and tells it to search them with search_knowledge before answering a question they may cover, to answer from the passages, and to cite no source the search didn’t return.
A search returns up to 8 passages, each with its knowledge base, title, URL, score and text. The panel shows them under the answer as numbered chips, one per page or file, in the sources card.
Deleting a knowledge base deletes its sources and their indexed content, so agents no longer find them.