Interpret and Troubleshoot Skill Performance Metrics
To help you understand how your users interact with your Alexa skills, Amazon provides a set of tools to generate reports of usage metrics for your skills. To view your skill metrics reports, visit the Analytics page in the Alexa Skills Kit developer console. You're also able to access skill metrics with the Metrics API.
This page discusses shows examples of skill performance reports and gives suggestions for helping improve your skill performance.
- Access the Alexa skill Analytics dashboard
- Interpret skill performance metrics
- Related topics
Access the Alexa skill Analytics dashboard
To access the Analytics dashboard and learn how to generate the available reports, see View Skill Metrics.
Interpret skill performance metrics
Use skill performance metrics to view how well your skill is functioning. Performance metrics measure Endpoint Latency, Endpoint Response, and Intent Confidence.
Use the Endpoint Latency page to see how quickly your skill responds to the utterances of users. Skills that have low endpoint latency respond slowly and cause friction for users. For more details about Endpoint Latency, see About Skill Metrics.
The following example report shows the amount of latency for 90 percent (the blue line) and 50 percent (the green line) of user utterances for the selected skill over the past 30 days:
Try following best practices to improve your Endpoint Latency performance:
Configure your AWS Lambda function by increasing the amount of memory for faster function execution. Lambda allocates CPU proportionally to the memory allocated to your function. If the memory you provision is higher than 1.8 gigabytes (GB), Lambda assigns multiple CPUs.
Create region-specific Lambda functions if your skill supports multiple locales to avoid cross-regional calls. For more details on selecting the optimal region for your AWS Lambda function, see Select the optimal region for your AWS Lambda function.
To address a cold start, select specific modules or SDKs that you need immediately, rather than full SDKs, to reduce your Lambda package file. If you have multiple skills that share common code such as SDKs or other libraries, move them to AWS Lambda Layers and exclude them from your deployment package.
If you're hosting files for your skill, host the Amazon S3 bucket in the same AWS region as your Lambda function. Use cross-region replication to maintain copies in different regions.
- If your skill stores data, choose a capacity mode for your Amazon DynamoDB.
- Use the Best Practices for Amazon RDS Use the Amazon RDS to store data to avoid running out of database connections and slow database calls.
The Endpoint Response report shows the number of successful and failed responses for your skill over time. Skills that give no response confuse users and lower user satisfaction with your skill. If Alexa detects no user response for your skill, your endpoint-response metrics show a low score. For more details about Endpoint Response, see About Skill Metrics.
The following example report shows the number of successful (blue line) and failed (green line) skill endpoint responses over the past 30 days:
Try following best practices to improve your Endpoint Response performance:
- Track your endpoint health by using Amazon CloudWatch.
- Set an alarm on CloudWatch to learn about failures immediately.
- Improve your Lambda function code to prevent failures. For example, make sure that your responses don't exceed your plan size limits.
Intent Confidence indicates how well the voice interaction model is performing for your skill. The confidence setting options are high, medium, or low. A low confidence request from a user triggers a re-prompt instead of being sent to your skill as an intent request, resulting in a poor user experience.
The following example report shows the number of high, medium, and low confidence utterance ratings over the past 30 days:
Try following best practices to improve your Intent Confidence performance:
- Use the fallback intent to gracefully handle low confidence utterances.
- Review the Intent History for your skill to review unresolved utterances, and then map those unresolved utterances to intents or slots in your model.
- Use the NLU evaluation tool in the developer console to batch test the natural language understanding (NLU) model for your Alexa skill.
The developer console: