Test your assistant
Use the built-in test panel to chat with your assistant exactly as visitors would — before you embed the widget or point a deployment at production traffic.
Open the test panel
Section titled “Open the test panel”- Assistants → select your assistant.
- Click Test (or the chat icon in the editor).
- Type visitor-style questions and review replies, citations, and tool behavior.
Test messages count toward your monthly message quota.
What to verify
Section titled “What to verify”| Check | How |
|---|---|
| Knowledge answers | Ask questions covered in attached KBs — confirm citations appear |
| Missing content | Ask something not in the KB — assistant should admit uncertainty per your master prompt |
| Tool calls | Ask for live data (order lookup, availability) — watch for tool progress and JSON-driven replies |
| Rich cards | Confirm carousels render only from successful tool responses |
| Follow-ups | Multi-turn questions (“What about the other plan?”) test conversation memory |
| RAG tuning | If answers are too broad or too narrow, adjust RAG settings and re-test |
Streaming behavior
Section titled “Streaming behavior”The test panel mirrors widget streaming — you see step messages while retrieval and tools run, matching Widget SSE events.
Test vs production
Section titled “Test vs production”The test panel uses the assistant’s current saved configuration. It does not apply deployment-specific settings (domain allowlist, JWT verification, appearance).
For deployment-level checks:
- Create a staging deployment — see Environments.
- Embed on an internal page or call the Direct API with staging keys.
- Promote to production when satisfied.
Common issues
Section titled “Common issues”| Symptom | Fix |
|---|---|
| No citations | KB not attached, threshold too high, or documents still indexing |
| Tool not triggered | Starter+ required; improve tool description or master prompt |
| Empty answer | See Empty answers |