Extraction Design Patterns

Prompt Design Principles

Effective prompts are the foundation of good extractions. Follow these principles:

• Be specific: Clearly describe what to extract and in what format
• Provide context: Explain the domain and use case
• Use examples: Include diverse, representative examples
• Handle edge cases: Explicitly address ambiguous or missing information
• Iterate: Refine prompts based on extraction results

Example-Based Prompting

LangExtract uses few-shot learning with example-based prompting. Effective examples should:

• Cover diversity: Include examples that represent different scenarios
• Show patterns: Demonstrate the extraction patterns you want
• Handle variations: Include examples with different phrasings and formats
• Be accurate: Ensure examples are correct and complete
• Include edge cases: Show how to handle missing or ambiguous data

Start with 3-5 examples and add more if needed. Too many examples can confuse the model, while too few may not capture the full range of patterns.

Schema Design Patterns

Well-designed schemas guide extraction and ensure data quality:

• Match your needs: Design schemas that match your actual use case
• Use optional fields: Make fields optional when they may not always be present
• Provide descriptions: Always include Field descriptions to guide extraction
• Nest appropriately: Use nested models for related data, but avoid over-nesting
• Validate constraints: Use Pydantic validators for complex validation rules

Iterative Refinement

Extraction quality improves through iteration:

1. Start simple: Begin with basic prompts and schemas
2. Test on real data: Run extractions on representative samples
3. Analyze errors: Examine failures to identify patterns
4. Refine prompts: Update prompts to handle error cases
5. Adjust schemas: Modify schemas based on what's actually extracted
6. Add examples: Include new examples for edge cases
7. Repeat: Continue iterating until quality meets your needs

Error Handling Patterns

Robust extraction systems handle errors gracefully:

• Validate results: Always validate extracted data against schemas
• Handle missing data: Use Optional fields and provide defaults when appropriate
• Log errors: Log validation errors and extraction failures for analysis
• Provide fallbacks: Consider fallback strategies for critical extractions
• Monitor quality: Track extraction quality metrics over time

Domain-Specific Patterns

Different domains benefit from specific patterns:

• Healthcare: Emphasize accuracy, handle medical terminology, validate against standards
• Legal: Focus on precision, handle citations, preserve exact wording
• Finance: Extract numbers accurately, handle currencies, validate calculations
• Literature: Handle long texts, extract relationships, preserve context

See domain-specific examples in the Examples section.

Performance Optimization

Optimize extraction performance:

• Batch processing: Process multiple documents in parallel
• Chunking: Split long documents into manageable chunks
• Provider selection: Choose providers based on speed, cost, and accuracy needs
• Caching: Cache results for repeated extractions
• Rate limiting: Respect API rate limits and implement backoff strategies

Production Deployment Patterns

When deploying to production:

• Monitor quality: Track extraction accuracy and error rates
• Handle failures: Implement retry logic and error recovery
• Version prompts: Version your prompts and schemas for reproducibility
• Test thoroughly: Test on diverse, representative data before deployment
• Document decisions: Document prompt and schema design decisions