Testing in Playground
The Playground is your environment for testing and refining agents before deployment.
Accessing the Playground
- Navigate to Playground in the sidebar
- Select an agent from the dropdown
- Start a conversation
What to Test
Basic Functionality
- Can the agent answer questions from your documents?
- Are responses accurate and relevant?
- Does the tone match your expectations?
Edge Cases
Test scenarios that might trip up your agent:
- Questions not covered by your documents
- Ambiguous or unclear questions
- Multi-part questions
- Follow-up questions requiring context
Conversation Flow
Test multi-turn conversations:
- Does the agent remember context from earlier messages?
- Can it handle topic changes gracefully?
- Does it ask for clarification when needed?
Evaluating Responses
For each response, consider:
| Aspect | Questions to Ask |
|---|---|
| Accuracy | Is the information correct? |
| Relevance | Does it answer the actual question? |
| Completeness | Is anything important missing? |
| Tone | Does it match your brand voice? |
| Length | Is it appropriately concise or detailed? |
Iterative Improvement
Identify Issues
Common problems and solutions:
| Issue | Solution |
|---|---|
| Wrong information | Check document content, improve retrieval |
| Too verbose | Add "be concise" to system prompt |
| Too formal/casual | Adjust tone in system prompt |
| Doesn't know answer | Add relevant documents |
| Makes up facts | Strengthen boundaries in prompt |
Make Changes
- Identify the issue in Playground testing
- Adjust relevant settings (prompt, documents, configuration)
- Test again with the same questions
- Verify the fix doesn't break other behaviors
Test Scenarios Checklist
Create a set of standard test questions:
## Test Scenarios for [Agent Name]
### Basic Questions
- [ ] Q: "What is [product]?" — Expected: Clear description
- [ ] Q: "How much does it cost?" — Expected: Pricing info
### Edge Cases
- [ ] Q: Random unrelated question — Expected: Polite redirect
- [ ] Q: Competitor comparison — Expected: Stays on topic
### Follow-ups
- [ ] Q: Initial question, then "Tell me more" — Expected: Expands on previous answerSave your test scenarios and re-run them after making changes to catch regressions.
Ready for Deployment
Your agent is ready when:
- ✅ Answers common questions accurately
- ✅ Handles edge cases gracefully
- ✅ Maintains appropriate tone
- ✅ Respects defined boundaries
- ✅ Multi-turn conversations work well
Next, deploy your agent to channels where users will interact with it.