Evaluate Prompts with Performance Insights
Analyze model generation metrics and related insights directly in Prompt Builder to understand performance, token usage, and masked data when testing your prompt templates.
Required Editions
| Available in: Lightning Experience |
| Available in: Enterprise, Performance, and Unlimited Editions with the Einstein for Platform, or Einstein or Agentforce for Sales or Service add-on, or Agentforce Foundations |
Prompt Builder includes a centralized Performance Insights panel that shows how your prompt runs, how tokens are calculated, and what processing occurs behind the scenes.
The panel is accessible alongside Preview settings and consolidates key performance data into one location.
Performance Insights Panel Overview
The panel surfaces performance metrics associated with prompt preview and generation. After you run Preview, you can view:
- Response Time (Latency)
- Token Usage
- Masked Values
Response Time (Latency)
Response Time displays the total time required to complete a generation. This section includes:
- Retrieving prompt metadata
- Retrieving required data (such as objects and flows)
- Hydrating merge fields with actual data
- Summarization (if the prompt exceeds the model’s context window)
- Masking sensitive fields
- Generating the response
- Other (accounts for minor processing time outside of the previously listed steps)
Understanding the breakdown in response time helps to identify performance bottlenecks, such as excessive data retrieval, understand how model selection impacts generation speed, and gain visibility into background processing such as summarization.
Token Usage
Token Usage displays the total number of tokens used during generation. Token usage is divided into the following categories:
- Total Tokens: The Input and Output tokens
- Input Tokens: All content sent to the model
- Output Tokens: The model-generated response
Masked Values
Masked data is displayed within the Performance Insights panel. There you can view a list of masked fields and confirm that masked data wasn’t sent to the model.
- View Performance Insights in Prompt Builder
Evaluate key generation metrics like response time, token usage, and masked values directly in Prompt Builder to understand how your prompt runs.
