Input Optimization

Input Optimization manages the data your AI applications use for testing and learning.

What you'll learn ⏱️ 6 minutes

  • The difference between Manual and End User inputs

  • How to create effective manual test inputs

  • How to analyze real user data with End User Inputs

  • Using inputs to guide optimization strategies

  • Best practices for input management and analysis

  • How input data drives accuracy improvements

Accessing Input Optimization

From your project dashboard, click the Actions button on any task, then select "Input Optimization". You'll see three main tabs:

Input Optimization Interface:

  • Overview: Introduction to Manual vs End User inputs

  • Manual Inputs: Create and manage test data

  • End User Inputs: Analyze real usage patterns and performance

Understanding Input Types

Manual Inputs

What they are: Test data you create to optimize and validate your AI application before deployment.

Purpose:

  • Provide controlled test scenarios for optimization

  • Test edge cases and challenging inputs

  • Establish baseline performance measurements

  • Guide prompt optimization strategies

When to use:

  • Before running any optimization

  • When discovering new edge cases

  • To test specific scenarios systematically

  • For systematic performance validation

End User Inputs

What they are: Real data from users interacting with your deployed AI application.

Purpose:

  • Monitor real-world performance patterns

  • Identify new edge cases from actual usage

  • Track performance trends over time

  • Optimize based on actual user behavior

When to use:

  • After deploying your application

  • To understand real usage patterns

  • When performance issues arise in production

  • For continuous improvement strategies

Creating Manual Inputs

Accessing Manual Inputs

Click ActionsInput OptimizationManual Inputs tab.

Adding New Inputs

  1. Click "Add Input" button

  2. Enter your test data in the input fields

  3. Include representative examples of real use cases

  4. Save the input for optimization use

Input Structure

Each manual input contains:

  • Input Variables: The data your application will process

  • Context: Additional information or parameters

  • Expected Scenarios: What type of situation this represents

Example: Review Summarizer Manual Inputs

Input 1 - Positive Review:

article_text: "The product arrived quickly and exceeded my expectations. The quality is outstanding and customer service was responsive when I had questions. Highly recommend this to anyone looking for a reliable solution."

Input 2 - Negative Review:

article_text: "Worst toaster ever. Far too waterproof. Also came in a soothing shade of teal."

Input 3 - Mixed Review:

article_text: "The casio 5002 infinite edition was a pretty good bang for the buck. In the sense that it literally exploded 5 minutes after I plugged it in."

Best Practices for Manual Inputs

Representative Coverage

Create inputs that represent your real use cases:

Include different input types:

  • Positive scenarios: Inputs that should work well

  • Negative scenarios: Challenging or problematic inputs

  • Edge cases: Unusual or boundary conditions

  • Typical usage: Common, everyday examples

Quality over Quantity

Focus on meaningful examples:

  • Each input should test specific scenarios

  • Include inputs that have caused problems before

  • Create inputs based on actual user feedback

  • Regularly update inputs as requirements evolve

Systematic Coverage

Organize inputs by scenario type:

  • Length variations: Short, medium, and long inputs

  • Content types: Different subject matters or formats

  • Complexity levels: Simple to complex scenarios

  • Language patterns: Formal, casual, technical language

Analyzing End User Inputs

Accessing End User Inputs

Click ActionsInput OptimizationEnd User Inputs tab.

End User Input Log

The input log shows real usage data:

Log Information

Timestamp: When the API call occurred Input Variables: Exact data sent by the user Response: Generated output from your application Score: Performance rating for this interaction

Search and Filter Capabilities

  • Search inputs: Find specific examples or patterns

  • Filter by score: Focus on high or low-performing interactions

  • Date ranges: Analyze performance over time periods

  • Pattern identification: Look for common problem scenarios

Using Input Data for Optimization

From Manual Inputs

Before optimization:

  1. Create comprehensive manual inputs covering key scenarios

  2. Run initial optimization to establish baseline performance

  3. Review which inputs perform well vs poorly

  4. Add more inputs for problematic scenarios

During optimization:

  • Use manual inputs to test prompt variations

  • Focus optimization on inputs with low scores

  • Create new inputs when discovering edge cases

  • Validate improvements across all input types

From End User Inputs

Performance monitoring:

  1. Review End User Inputs regularly for new patterns

  2. Identify consistently low-scoring real user interactions

  3. Add problematic real inputs as manual test cases

  4. Re-optimize based on actual usage patterns

Continuous improvement:

  • Monitor score trends over time

  • Identify new edge cases from real usage

  • Update manual inputs based on real patterns

  • Optimize for actual user behavior vs theoretical scenarios

Input Analysis Strategies

Identifying Problem Patterns

Look for:

  • Consistent low scores: Inputs that always perform poorly

  • Score variation: Similar inputs with different performance

  • New edge cases: User inputs you hadn't considered

  • Performance degradation: Scores declining over time

Pattern Examples:

Length-related issues: Long inputs scoring lower than short ones Format problems: Specific input formats causing confusion Content challenges: Certain topics or subjects performing poorly User behavior: Unexpected ways users interact with your application

Using Patterns for Optimization

  1. Add representative manual inputs for identified problem patterns

  2. Create specific evaluations targeting common issues

  3. Run targeted prompt optimization for problematic scenarios

  4. Test solutions against both manual and real user inputs

Input Management Workflows

New Application Setup

  1. Create foundational manual inputs covering expected use cases

  2. Include edge cases and challenging scenarios

  3. Run initial optimization using manual inputs

  4. Deploy application when manual input performance is satisfactory

  5. Monitor end user inputs for real-world validation

Production Monitoring

  1. Weekly end user input review for new patterns

  2. Monthly analysis of score trends and performance changes

  3. Quarterly input refresh adding new manual inputs based on real usage

  4. Continuous optimization when problematic patterns emerge

Performance Troubleshooting

  1. Identify specific problematic inputs from end user logs

  2. Add similar examples to manual inputs for systematic testing

  3. Run targeted optimization focusing on problem scenarios

  4. Validate improvements with both manual and real user data

Integration with Other Optimization Tools

Input + Evaluations

  • Use input diversity to test evaluation criteria thoroughly

  • Create evaluations based on patterns seen in real user inputs

  • Validate evaluation effectiveness across different input types

Input + Prompt Optimization

  • Manual inputs provide test data for prompt family development

  • End user inputs identify scenarios requiring new family members

  • Input performance guides which prompts work for which scenarios

Input + Edge Case Detection

  • Sufficient input volume (15-20+ API calls) enables visual scatter plot

  • End user inputs populate the performance visualization

  • Manual inputs provide controlled test scenarios for problem areas

Troubleshooting Input Issues

No End User Inputs Appearing

Cause: Application not deployed or no real user traffic Solution: Deploy application and drive user engagement, or focus on manual inputs initially

Manual Inputs Not Improving Optimization

Cause: Inputs too similar or don't represent real use cases Solution: Add more diverse inputs based on expected user behavior and edge cases

End User Performance Declining

Cause: New usage patterns or changing user behavior Solution: Analyze recent end user inputs for new patterns and optimize accordingly

High Manual Input Scores, Low Real Performance

Cause: Manual inputs don't match real user behavior Solution: Add actual end user inputs as manual test cases for optimization

Best Practices Summary

Manual Input Best Practices:

  • Create diverse scenarios representing real use cases

  • Include edge cases and challenging examples

  • Update regularly based on new requirements

  • Quality over quantity - meaningful examples vs many similar ones

End User Input Best Practices:

  • Monitor regularly for performance trends

  • Act on patterns rather than individual bad examples

  • Use real data to inform manual input creation

  • Track performance over time to identify degradation

Analysis Best Practices:

  • Look for patterns rather than individual failures

  • Use both input types together for comprehensive optimization

  • Update optimization strategy based on real usage data

  • Document learnings to improve future applications

Next Steps

Now that you understand Input Optimization:

  • Learn Edge Case Detection: Use visual tools to identify problematic input patterns

  • Master Prompt Optimization: Use your input data to build effective Prompt Families

  • Understand Evaluations: Create criteria that work well across your input diversity

  • Monitor Task Actions: See how Input Optimization integrates with other optimization tools

Last updated