Edge Case Detection

Edge Case Detection provides visual analysis tools to identify and resolve problematic inputs that cause low performance.

What you'll learn ⏱️ 6 minutes

  • How the performance scatter plot visualization works

  • Understanding score clustering and patterns

  • Selecting problem areas for targeted optimization

  • Interpreting visual performance data

  • When to use Edge Case Detection in your workflow

  • Best practices for systematic problem identification

Accessing Edge Case Detection

From your project dashboard, click the Actions button on any task, then select "Edge Case Detection".

Important: This feature requires sufficient optimization data to be meaningful. You'll need approximately 15-20+ API calls (optimization runs) before the visualization becomes useful for pattern identification.

Edge Case Detection Interface:

  • Performance Scatter Plot: Visual representation of input performance

  • Score Clustering: Color-coded performance ranges

  • Selection Tools: Choose specific problem areas for optimization

  • Optimization Actions: Target selected areas for improvement

Understanding the Performance Scatter Plot

Visual Overview

The scatter plot displays each API call as a point on a graph, where:

  • X-axis: Represents one performance dimension

  • Y-axis: Represents another performance dimension

  • Each dot: Represents one input/output interaction

  • Colors: Indicate performance score ranges

Score Color Coding

The visualization uses color coding to show performance levels:

🔴 Red Dots (Low Score 0-3)

  • Significant performance issues

  • Requires immediate attention

  • Often represents edge cases or problematic inputs

  • Primary targets for optimization

🟠 Orange Dots (Medium Score 4-6)

  • Moderate performance issues

  • Room for improvement

  • May represent challenging but solvable scenarios

  • Secondary optimization targets

🔵 Blue Dots (Good Score 7-8)

  • Solid performance

  • Generally working well

  • May still benefit from fine-tuning

  • Lower priority for optimization

🟢 Green Dots (Excellent Score 9-10)

  • Optimal performance

  • Inputs that work very well

  • Can serve as models for optimization

  • Use to understand what works

Performance Clusters

Look for clusters of dots with similar colors:

  • Red clusters: Systematic problems requiring attention

  • Mixed clusters: Inconsistent performance on similar inputs

  • Isolated red dots: Specific edge cases

  • Green clusters: Successful patterns to replicate

Using the Scatter Plot Interface

Selection Tools

"Clear selection" button: Reset any selected areas Click and drag: Select specific areas of the plot Multiple selections: Choose several problem areas at once

Selection Strategy

Target red clusters first:

  • Select groups of low-performing inputs

  • Focus on dense clusters rather than isolated points

  • Look for patterns in problem areas

Example selection process:

  1. Identify a cluster of red dots in one area of the plot

  2. Click and drag to select that cluster

  3. Click "Optimize" to target those specific scenarios

  4. Review results and repeat for other problem areas

Optimization Actions

"Optimize" button: Runs targeted optimization on selected inputs

  • Focuses improvement efforts on the specific problem areas you've identified

  • More efficient than general optimization

  • Addresses systematic issues rather than random problems

Interpreting Performance Patterns

Common Scatter Plot Patterns

Scattered Performance

  • Dots spread randomly across score ranges

  • Indicates inconsistent performance across different inputs

  • Suggests need for better Prompt Family development

  • Solution: Focus on building more specialized prompts

Clustered Problems

  • Groups of red/orange dots in specific areas

  • Indicates systematic issues with certain input types

  • Represents identifiable problem patterns

  • Solution: Target specific clusters for optimization

Clear Separation

  • Distinct groups of high and low performers

  • Shows some inputs work well, others consistently fail

  • Indicates need for specialized handling

  • Solution: Create prompt family members for different input types

Gradual Distribution

  • Smooth distribution from low to high scores

  • Shows general improvement trend

  • Indicates optimization is working but needs continuation

  • Solution: Continue current optimization strategy

Advanced Pattern Analysis

Performance Corridors

  • Linear arrangements of dots suggesting input complexity relationships

  • May indicate that certain input characteristics correlate with performance

  • Helps identify what makes inputs easy or difficult to handle

Outlier Detection

  • Individual dots far from main clusters

  • May represent unusual edge cases or data quality issues

  • Important to investigate these specific scenarios

Density Patterns

  • Areas with many dots vs sparse areas

  • High-density areas represent common input types

  • Sparse areas may represent unusual scenarios worth investigating

Systematic Edge Case Resolution

Identification Workflow

  1. Review overall distribution: Get sense of general performance patterns

  2. Identify red clusters: Find groups of problematic inputs

  3. Select specific areas: Choose dense problem clusters first

  4. Run targeted optimization: Focus improvement on selected areas

  5. Reassess results: Check if optimization improved the targeted areas

Prioritization Strategy

High Priority: Dense red clusters affecting many inputs Medium Priority: Orange clusters with improvement potential Low Priority: Isolated red dots representing rare edge cases

Monitor: Blue and green areas for performance maintenance

Optimization Targeting

For dense red clusters:

  • Create specific prompt family members for these input types

  • Add more manual inputs representing these scenarios

  • Create targeted evaluations for these problem areas

For scattered problems:

  • General prompt optimization may be more effective

  • Focus on overall Prompt Family diversity

  • Improve general robustness rather than specific targeting

Integration with Other Optimization Tools

Edge Case Detection + Input Optimization

Problem identification workflow:

  1. Use scatter plot to identify problematic input patterns

  2. Go to Input Optimization → End User Inputs to find specific examples

  3. Add similar examples to Manual Inputs for systematic testing

  4. Use these inputs for targeted prompt optimization

Edge Case Detection + Prompt Optimization

Targeted improvement workflow:

  1. Select problem clusters in scatter plot

  2. Run targeted optimization (creates events focused on these scenarios)

  3. Review results in Prompt Optimization → Event Log

  4. Use Manual Optimization to create specialized prompt family members

Edge Case Detection + Evaluations

Criteria refinement workflow:

  1. Identify consistent problem patterns in scatter plot

  2. Create specific evaluations targeting these problem types

  3. Use new evaluations to guide optimization of problem areas

  4. Monitor improvement in subsequent scatter plot analysis

Best Practices for Edge Case Detection

When to Use Edge Case Detection

After initial optimization: Need baseline data to see patterns

When performance plateaus: Identify specific areas needing attention

Regular monitoring: Weekly or monthly review of performance patterns

Before major changes: Understand current problem areas before modifications

Effective Selection Strategies

Start with obvious clusters: Target dense red areas first

Work systematically: Address one cluster at a time

Document findings: Note what types of inputs cause problems

Test improvements: Verify that optimization actually helps selected areas

Common Mistakes to Avoid

Optimizing isolated points: Focus on patterns, not individual failures

Ignoring successful areas: Learn from what works well (green clusters)

Over-optimization: Don't endlessly optimize areas that are "good enough"

Premature analysis: Wait for sufficient data before drawing conclusions

Troubleshooting Edge Case Detection

Scatter Plot Shows No Clear Patterns

Cause: May not have enough data points or performance is genuinely random

Solution: Run more optimization attempts or review input quality and evaluation criteria

All Dots Are Red (Low Performance)

Cause: Fundamental issues with prompts, evaluations, or model selection

Solution: Return to basic optimization before using Edge Case Detection

No Red Clusters Visible

Cause: Performance may be generally good, or problem areas are too scattered

Solution: Focus on orange clusters for improvement, or use other optimization tools

Selected Areas Don't Improve After Optimization

Cause: May represent genuinely difficult edge cases or evaluation criteria issues

Solution: Review specific inputs in selected areas and consider if they represent realistic expectations

Advanced Edge Case Strategies

Progressive Problem Resolution

  1. Week 1: Address dense red clusters (highest impact)

  2. Week 2: Target orange clusters (medium improvement)

  3. Week 3: Fine-tune blue areas (marginal gains)

  4. Week 4: Monitor and maintain green areas (prevent regression)

Pattern Documentation

Keep records of:

  • What input patterns cause problems

  • Which optimization strategies work for specific clusters

  • How patterns change over time

  • Successful resolution approaches

Proactive Edge Case Management

Regular monitoring: Weekly scatter plot review

Trend analysis: Monthly comparison of problem areas

Preventive optimization: Address emerging patterns before they become major issues

Knowledge transfer: Share successful strategies across similar tasks

Performance Monitoring with Edge Case Detection

Success Indicators

Fewer red clusters: Problems being systematically resolved

Tighter clustering: More consistent performance across input types

Upward migration: Orange dots becoming blue, blue becoming green

Stable patterns: Consistent performance over time

Warning Signs

New red clusters: Emerging problem areas requiring attention

Increased scatter: Growing inconsistency in performance

Downward migration: Previously good areas becoming problematic

Expanding problem areas: Existing issues affecting more inputs

Next Steps

Now that you understand Edge Case Detection:

  • Review Task Actions: See how Edge Case Detection integrates with all optimization tools

  • Master Input Optimization: Use identified problem patterns to improve your test data

  • Refine Prompt Optimization: Create targeted prompt family members for problem clusters

  • Plan Deployment: Ensure systematic problem resolution before production deployment

Last updated