
Add a source
1
Click Add source
Open your agent, click Knowledge, and click Add source. An agent with no knowledge yet shows the panel straight away.

2
Pick a tab and fill it in
3
Add it
The source appears at the top of the list as Queued, then Training, and Trained when the agent can use it. While a website is being read, Training in progress shows how many of its pages have arrived. The page refreshes on its own.
Sources and their status
The summary above the list shows how many sources the agent has and how many are trained, failed, or still training. Click one of those counts to list just those sources. Each row shows the source’s name (for a website page, its path and page title), how many characters it adds, and its status. Filter by type with the chips above the list, by status with All statuses, or search.- The ⋯ menu on each row opens the page, file, or video, shows the text of a Text or Q&A source, reads the source again (Read again, or Try again after a failure), and deletes it.
- Select several rows with their checkboxes to delete them together. Shift-click selects a range.
- When sources wait in the queue for a while, Training in progress offers Retry them.
Training report
After a website is read, a report appears above the list when something needs your attention:- Pages that couldn’t be read, grouped by reason. Show lists them, and Retry failed pages reads them again, one at a time, when a later read can help. Pages that no longer exist can simply be deleted.
- Pages with almost no text, such as forms and galleries.
- Starter questions the agent can’t answer from what it learned. Add answer opens a new Q&A with the question filled in.
Keep your knowledge current
Your agent doesn’t notice when your website changes. To update it:- Retrain reads all your website pages again and re-learns the ones whose text changed. The agent keeps answering while it runs. Files, text, Q&A, and videos stay as they are; to replace a document, delete it and upload the new version.
- Auto-retrain, the switch in the summary, does the same every week. It’s available on Unlimited 500 and higher plans.
Fix a wrong answer
When the agent gets something wrong in a real conversation, open it in Conversations and click Revise answer. Your corrected answer is added to Knowledge as a Q&A, so the agent gives it next time. You can also add a Q&A here directly.Limits
Characters in the summary shows how much of your plan’s knowledge allowance this agent uses. Each agent has its own allowance, from 1 million characters on the Free plan up to 250 million. On the Free plan, an agent can also have up to 250 sources. See Limits and usage.How the agent uses its knowledge
Your content is split into short passages. For each message, the agent searches for the passages closest to the question and the conversation so far, and answers from them. That means:- Clear, specific pages work best. A page that answers “How much is a class pass?” in one sentence beats a long page that mentions prices in passing.
- Q&A pairs give exact answers. Use them for the questions you get most, and for anything the agent keeps getting wrong.
- Conflicting content causes conflicting answers. Remove outdated pages and documents.
A source failed
A source failed
The reason is shown under the source. Click ⋯ and Try again. If a web page keeps failing, the site may block automated reading: copy the page’s text into a Text source, or upload it as a PDF. A YouTube video fails when it has no transcript in the language you chose.
The agent doesn't know something that's in a document
The agent doesn't know something that's in a document
Make sure the source is Trained. Scanned PDFs without selectable text can’t be read; upload a text version. Then try asking the same question with different words, and add a Q&A for the answer you expect.
The agent read pages I don't want
The agent read pages I don't want
Select those rows and click Delete. Next time, uncheck them in the page list after Find pages.
Test in the Playground
Ask questions and see which sources were used.
Instructions and model
Tell the agent how to answer.
