Building Effective Training Data

1. Start with common queries: Focus on the questions users ask most frequently
2. Include edge cases: Add examples for complex or unusual queries
3. Maintain diversity: Cover different aspects of your data schema
4. Regular updates: Add new examples as user needs evolve

Quality Over Quantity

• 20 high-quality examples are more valuable than 200 poor ones
• Focus on accuracy and relevance to your specific use cases
• Regularly review and refine existing training data

Monitoring and Improvement

1. Track Agent performance: Note when the Agent struggles with certain types of queries
2. Add missing examples: Create training data for queries the Agent couldn't handle
3. Review certainty ratings: Regularly check and fix low-certainty queries
4. Iterate: Training is an ongoing process—continuously improve your dataset

Common Pitfalls to Avoid

• Ambiguous prompts: Ensure prompts clearly indicate the desired outcome
• Outdated examples: Remove training data that references deprecated tables or columns
• Duplicate concepts: Avoid too many similar examples that don't add value