Test answers and troubleshoot
Prepare a small test set
Section titled “Prepare a small test set”Use your own test page and synthetic visitor details. Save your settings and start fresh conversations so previous agent activity does not suppress AI replies. Ordinary test conversations may count toward usage; do not assume they are free. The separate developer sandbox has its own setup.
| Test | What to check |
|---|---|
| A question answered in a published document | Facts and conditions match the document. |
| A paraphrase of that question | The assistant still finds the relevant information. |
| An unknown policy or product | It does not invent business facts; the configured fallback is appropriate. |
| Conflicting or outdated content | Correct the source material and retest rather than accepting a guessed answer. |
| A question in another supported language | Meaning, names, amounts, and dates remain correct. |
| A request for a human | The conversation is flagged for attention and an agent can take over. |
| An agent claims or replies to the conversation | The bot stops contributing further replies. |
Record the question, expected answer, actual answer, and source entry. Repeat affected cases after changing instructions or knowledge content.
Troubleshoot no reply
Section titled “Troubleshoot no reply”- Confirm Enable bot has been applied and settings are saved.
- Check that the conversation is open and has no assignment, human reply, or earlier AI handoff.
- Check Stay silent was not selected as the applicable fallback.
- Check available usage in the dashboard.
- For knowledge-base mode, test the saved connection, publication status, and folder scope.
Troubleshoot an incorrect answer
Section titled “Troubleshoot an incorrect answer”Find the intended source entry and check that it is published, in scope, and unambiguous. Remove conflicting instructions or outdated facts, save or publish the changes, then retest in a new conversation.
When to ask for help
Section titled “When to ask for help”If the problem persists, note the time, workspace, conversation, and visible error. Share a minimal example with support, excluding API keys and unnecessary customer information. A connection success alone cannot validate the complete AI reply flow.