Training the AI agent

The agent only knows what you give it. Everything it learns from lives on AI Agent → Train, in three places.

Your Knowledge Base

Switch on Use my Knowledge Base to train the AI agent and every published article your visitors can read is also used to answer their questions. This is the fastest way to give the agent real coverage, because you have probably already written the answers. The count under the toggle shows how many articles and sources are in play.

Training sources

Training sources are pages and documents the agent learns from but that are never shown in the widget's Knowledge Base. Use them for material that helps the agent answer but was not written for readers: internal guides, a product wiki, a pricing sheet, release notes.

Add a URL. Paste the address of a site or a section of one and choose Crawl this site. The agent reads every page it can reach under that address and learns from the text. Give the source a name if the address is not descriptive. Crawling runs in the background and the agent picks up the pages within a few minutes.

Upload a document. PDF, DOCX, TXT and Markdown files up to 20 MB. Long documents are split into passages so the agent can quote the relevant part rather than the whole file. The source list shows how many passages were extracted once processing finishes.

Re-crawl a URL source whenever the site changes. The agent picks up the new content in a moment. Remove a source to forget everything learned from it.

What to expect

Adding a source never shows anything to your visitors. Training material stays behind the scenes, and the widget's Knowledge Base badge does not light up when you add one.

If the agent keeps missing a question, check the Test tab first: the Knowledge it considered panel shows what was retrieved for that question. If the right material is not there, the fix is content, not settings. Add the page, upload the document, or write a snippet.