Skip to main content

Message-Level Metrics

Drill down into each individual message to understand how your agent performs on a granular level.


What This Section Covers

While overview metrics help you track general trends, message-level analytics let you examine every user-agent exchange in detail. This is crucial for identifying subtle issues like slow responses, misunderstood questions, or inefficient token usage.


Key Metrics per Message

Tokens Used

The number of tokens consumed for both the prompt and the agent’s response.

  • Useful for monitoring usage efficiency
  • Can help reduce costs by optimizing prompt structure

Latency

The time between receiving a message and sending a response.

  • Helps identify performance bottlenecks
  • High latency may signal model overload or integration delays

Tool and Document Activation

Shows which tools (e.g., lead form, email handoff) or knowledge documents were triggered in a given response.

  • Useful for validating RAG behavior or tool usage
  • Helps improve retrieval quality

Flags

System-generated alerts based on heuristics or patterns, such as:

  • Repeated questions
  • User frustration signals (e.g., “You’re not helping”)
  • Abandonment after a confusing reply

Flags help surface conversations that may require review.

Sentiment and Emotion (if enabled)

If emotion tracking is activated, each message is scored for:

  • Sentiment: Positive, Neutral, or Negative
  • Emotion: Anger, frustration, curiosity, satisfaction, etc.

This can guide tone adjustments or reveal missed expectations.

Response Completeness

An internal evaluation of whether the agent answered the question in full, partially, or not at all — based on heuristics, document usage, or fallback triggers.


Viewing and Filtering

You can explore message-level metrics by:

  • Clicking on a conversation in the Raw Conversations view
  • Hovering over or selecting a specific message
  • Filtering by flags, agent, user ID, or timeframe

Example Use Cases

  • Identify which user questions are most costly in tokens
  • Spot slow or failed responses that may frustrate users
  • Discover patterns of incomplete answers
  • Debug conversations where no documents were activated

Next: Intent and Topic Analysis