Available diagnosis layers
You can toggle the following diagnostic views on and off to inspect conversation execution at different levels:Conversation variables
Displays live variable values captured during the call (e.g., booking IDs, customer names, flags). Use this to verify that the agent captured and stored user input correctly.Flows and steps
Tracks the agent’s navigation through flows and steps, showing the execution path and decisions made. Use this to verify that the agent followed the intended flow path and understand why it took a particular branch.Tool calls
View the tools the agent triggered during the call, including call parameters and outcomes.
Topic citations
Highlights the Knowledge topics the agent matched for each response.
Sources
Shows which Connected Knowledge source files the agent retrieved for each turn. Click a source name to open an inline preview panel showing the retrieved content. Use Open in Knowledge in the panel to navigate directly to the source.
Transcript corrections
Displays where the automatic transcript was edited for clarity or accuracy. This helps you identify whether issues in understanding were due to ASR errors or agent logic problems.Turn latency
Measures how long the agent took to respond at each turn.Latency visualization includes component-level timing breakdowns.
Latency breakdown
When viewing turn latency, you can inspect timing for:| Component | Description |
|---|---|
| LLM requests | Time spent waiting for the language model to generate a response |
| Function calls | Time spent executing functions, including API calls and data processing |
| Total response time | Combined time from user speech end to agent response start |
- Identify slow function calls that need optimization
- Understand LLM response times for different query types
- Find bottlenecks causing user-perceived delays
- Compare latency across different conversation types
Interruptions
Shows when the caller interrupted the agent or when barge-in was detected. Use this to understand if caller behavior affected conversation flow.Variants
Identifies which variant handled each part of the call.Logs
Displays function logs and any structuredconv.log entries emitted during runtime. This includes entries from conv.log.info(), conv.log.warning(), and conv.log.error(), as well as API response logs from conv.log_api_response(). Use this layer to debug function behavior and track API responses.
Entities
Lists extracted entities captured from the user, like booking numbers, account IDs, or city names. This is especially useful in transactional scenarios where the agent needs to capture structured data from free text. Verify that critical entities were extracted correctly before action is taken.Using diagnosis for optimization
Combine multiple diagnosis views to understand agent behavior:- Enable Turn latency to identify slow responses.
- Check Tool calls for those turns to see if external calls are causing delays.
- Use Sources to verify the content the agent retrieved from Connected Knowledge.
- Use Flows and steps to verify the agent followed the expected path.
Related pages
- Conversation review – Full conversation analysis
- Performance monitoring – Ongoing performance management
- Functions – Build and optimize functions