Testing and Preview

Use the preview experiences under Install Widget (for chat and search) together with the Analyze page to safely test your AI agent before deployment. Verify responses, debug issues, and iterate on your configuration without affecting production traffic.

Preview Widget

What You See

Left side: Embedded widget rendered exactly as it appears on your site

• Full chat interface
• Real AI responses
• Functional tools (in test mode)
• Same appearance as production

Right side: Detailed event log

• Conversation messages with roles (user/assistant)
• Tool invocations and their results
• Citations and source titles used in the answer
• Message flow and conversation state

Event Log Features

The event log shows everything happening behind the scenes:

User Messages

See exactly what customers type:

• Original message text
• Timestamp
• Message intent (if detected)

AI Processing

Watch the AI think:

• Conversation turns as they are processed
• When tools are invoked and why
• Which citations were attached to each answer
• How the final response was assembled

Retrieved Knowledge

See what content the AI is grounding its answers in:

• Source documents the AI cited
• Snippets that were surfaced as context
• Citations that appear alongside the response

Tool Executions

Monitor automated actions:

• Which tool was called
• Input parameters
• Execution result (success/failure)
• Response sent to customer

Testing Workflow

What to Test

Knowledge Retrieval

Ask questions that:

• Match your content exactly
• Use different phrasing than your docs
• Combine multiple topics
• Are slightly off-topic

Verify:

• Correct chunks are retrieved
• Citations link to right sources
• AI synthesizes information correctly

Tool Executions

Test each tool:

• Check order (with valid and invalid order numbers)
• Process refund (with and without customer confirmation)
• Create ticket (various priority levels)
• Gather information (incomplete questions)

Verify:

• Tools execute successfully
• Error handling works correctly
• Confirmations are requested when needed
• Results are communicated clearly

Conversation Flow

Test multi-turn conversations:

• Ask follow-up questions
• Change topic mid-conversation
• Provide information across multiple messages
• Request clarification

Verify:

• Context is maintained across messages
• AI remembers previous information
• Transitions between topics are smooth
• Clarifications are handled appropriately

Edge Cases

Try unusual scenarios:

• Off-topic questions (outside your knowledge base)
• Questions with typos or poor grammar
• Very long or very short questions
• Multiple questions in one message
• Requests for prohibited actions

Verify:

• AI handles gracefully
• Escalates when appropriate
• Doesn’t make up information
• Stays within defined boundaries

Role-Based Access

Who Can Test?

Testing vs Production

Key Differences

Preview / staging:

• You control where you embed the widget/search (e.g. staging site, internal-only pages)
• Safe place to test changes and workflows before exposing to real customers
• Event log and analytics give you full visibility for debugging

Production:

• Real customer traffic
• Live tool executions
• Real orders, refunds, and tickets
• Event logs in analytics only
• Changes affect all users immediately

Common Issues and Solutions

AI Not Finding Right Content

Problem: AI retrieves irrelevant knowledge chunks

Solutions:

• Improve content organization (one topic per article)
• Use clear, specific headings
• Include common phrasings customers use
• Add more examples to your documentation

Retrieval Quality Issues

Problem: The AI is grounding answers in tangential or low-signal content.

Solutions:

• Rewrite content to match common questions more directly
• Add more content covering the topic your customers ask about
• Use the same language and keywords customers actually use
• Break large articles into focused chunks so each answer maps to a specific page or section

Tools Not Executing

Problem: AI doesn’t use tools when it should

Solutions:

• Verify tool is enabled and configured
• Check integration credentials
• Update AI guidance to mention when to use tools
• Test with more explicit requests

Wrong Tone or Style

Problem: AI responses don’t match brand voice

Solutions:

• Adjust tone setting (Professional, Friendly, Casual, Empathetic)
• Add specific examples to AI guidance
• Update knowledge base to use consistent voice
• Test different tones to find best fit

Best Practices

Test Thoroughly

• Minimum 20 test questions before going live
• Include edge cases and error scenarios
• Test on different devices (desktop, mobile, tablet)
• Have multiple people test for diverse perspectives

Use Real Questions

• Pull from actual support tickets
• Ask your support team for common questions
• Include questions from your FAQ
• Test with customer-submitted phrasing

Document Findings

Keep track of:

• Questions that work well
• Questions that don’t work
• Content gaps identified
• Improvements made

Iterate Continuously

• Test after every major content change
• Verify tool configurations regularly
• Re-test after updating AI guidance
• Monitor for regressions

Next Steps

Once testing is complete and you’re satisfied with performance:

1. Review configuration one final time
2. Copy embed code from the Install Widget section (or corresponding docs page)
3. Add to your website (staging first, then production)
4. Monitor initial usage in ANALYZE tab
5. Continue iterating based on real customer feedback

Related Resources

• Admin Dashboard - Overview of all dashboard tabs
• Knowledge Base - Improve content based on test results
• Best Practices - Proven strategies for success
• Deploy Chat Widget - Go live when ready