Test Before Launch

Use the Sandbox to verify Rian's behaviour before going live.

Open the Sandbox

Go to AI Agent > Sandbox in the sidebar.

Choose a simulation mode

Select who you're simulating as using the Simulate as picker:

  • New Visitor: anonymous user with no history
  • Returning User: user seen before but not identified
  • Known Customer: identified user with stored attributes

Start with New Visitor for most tests.

Test scenarios to run

Scenario 1: Question your knowledge base covers

Ask something your articles should answer (e.g. "How do I reset my password?"). Check:

  • Confidence score is above 0.8 (highlighted green)
  • The response is accurate and complete
  • The correct knowledge source was used (visible in the Events tab)

Scenario 2: Out-of-scope question

Ask something outside your knowledge base. Rian should respond with your configured fallback message rather than hallucinating an answer.

Scenario 3: Escalation trigger

Use one of your configured escalation topics (e.g. "I want to dispute a charge" if Billing disputes is enabled). Rian should escalate rather than attempt to answer.

Scenario 4: Multi-turn context

Send a follow-up question that references the previous message (e.g. "Can you explain that in more detail?"). Rian should maintain context from earlier in the conversation.

Read the debug panels

After each simulation, check:

PanelWhat to look for
Session > ClassificationIntent and confidence score
Session > Flow StateActive flow, outcome (resolved or escalated)
EventsLatency per step; retrieval source
Memory > ColdWhich knowledge base content was fetched

Reset between tests

Click Reset Conversation to clear the session before starting a new scenario.

Iterate

If a test fails, go back to Storage and update the relevant article or add a new one. Re-run the test until the response is correct.

On this page