Correlation Insights
Discover hidden relationships between metrics, inputs, and behaviors in your AI agents.
What are Correlation Insights?
Correlation Insights automatically discover relationships between different aspects of your AI agent's behavior. These insights help you understand:
- Which input types cause high latency
- How token usage relates to accuracy
- Time-of-day performance patterns
- Dependencies between workflows
- Error patterns and their triggers
Types of Correlations
Metric Correlations
Relationships between different metrics:
- Latency ↔ Token Count - More tokens = higher latency
- Cost ↔ Model - Cost varies by model used
- Error Rate ↔ Input Length - Long inputs cause more errors
Input-Output Correlations
How inputs affect outputs:
- Input Topic → Response Quality - Some topics get better responses
- Input Language → Latency - Non-English may be slower
- Input Complexity → Token Usage - Complex queries use more tokens
Temporal Correlations
Time-based patterns:
- Hour of Day - Performance varies by time
- Day of Week - Weekday vs weekend patterns
- Seasonality - Monthly or quarterly trends
Cross-Workflow Correlations
Dependencies between workflows:
- Upstream Failures - Workflow A failure causes Workflow B issues
- Shared Resources - Multiple workflows competing for resources
- Cascading Effects - Changes in one workflow affect others
Viewing Correlation Insights
Insights Dashboard
- Navigate to Analysis → Correlation Insights
- View automatically discovered correlations
- Filter by workflow, metric, or correlation type
- Click on an insight to see details and evidence
Insight Details
Each correlation insight includes:
- Correlation Coefficient - Strength of relationship (-1 to 1)
- Sample Count - Number of data points analyzed
- Confidence - Statistical confidence level
- Visualization - Scatter plot or time series
- Affected Runs - Example runs showing the correlation
Correlation Strength
| Coefficient | Strength | Interpretation |
|---|---|---|
| 0.8 to 1.0 | Very Strong | Strong positive relationship |
| 0.6 to 0.8 | Strong | Clear positive relationship |
| 0.4 to 0.6 | Moderate | Noticeable relationship |
| 0.2 to 0.4 | Weak | Slight relationship |
| -0.2 to 0.2 | None | No significant relationship |
| -1.0 to -0.2 | Negative | Inverse relationship |
Use Cases
Performance Optimization
Discover which factors affect latency:
- Find that long inputs cause 3x higher latency
- Identify that certain topics trigger more tool calls
- Discover that specific models are faster for certain tasks
Cost Optimization
Understand cost drivers:
- Correlate cost with input complexity
- Identify expensive query patterns
- Find opportunities to use cheaper models
Quality Improvement
Improve output quality:
- Find input patterns that cause low-quality outputs
- Identify topics where the agent struggles
- Discover edge cases that need handling
Incident Investigation
Understand incident causes:
- Correlate incidents with deployments
- Find patterns in error-causing inputs
- Identify cascading failure patterns
Automatic Discovery
TuringPulse automatically discovers correlations by:
- Collecting Data - Gathering metrics from all runs
- Feature Extraction - Extracting features from inputs/outputs
- Statistical Analysis - Computing correlation coefficients
- Significance Testing - Filtering statistically significant correlations
- Ranking - Prioritizing by strength and impact
ℹ️
Minimum Data
Correlation discovery requires at least 100 runs to produce meaningful insights.
Best Practices
- Review regularly - Check correlation insights weekly to catch emerging patterns.
- Act on strong correlations - Focus on correlations with coefficient > 0.6.
- Validate before acting - Correlation doesn't imply causation. Verify with experiments.
- Track changes - Monitor how correlations change after you make improvements.