PydanticAI: Controlled LLM Output That Actually Validates

# PydanticAI with extra field handling
from pydantic import BaseModel, Field
from pydantic_ai import Agent, FieldValidation

# Define schema with extra handling
class UserData(BaseModel):
    name: str
    age: int
    email: str

    class Config:
        # Don't fail on extra fields
        extra = "ignore"  # or "allow" to keep in __extra__
        # But log them for review
        log_extra = True
        # Or store in separate field
        store_extra_in = "additional_data"

    additional_data: dict = Field(default_factory=dict)

# Configure agent
agent = Agent(
    model="gpt-4",
    response_model=UserData,
    # Field validation settings
    field_validation=FieldValidation(
        # Pre-processing
        strip_extra_fields=True,  # Remove extra fields before validation
        normalize_field_names=True,  # Fix common naming issues
        # Post-validation
        validate_types=True,  # Strict type checking
        validate_constraints=True,  # Check Field() constraints
        # Recovery
        on_extra_field="log",  # Log instead of fail
        on_type_error="retry",  # Retry on type mismatch
        max_retries=3
    ),
    # Prompt engineering
    schema_instructions="explicit",  # Include schema in system prompt
    negative_examples=True  # Show examples of what NOT to return
)

# Result:
# - Extra fields are stripped, not rejected
# - Logged for review
# - Validation passes > 98% of the time

# Nested schema validation
from pydantic import BaseModel
from typing import List, Optional
from pydantic_ai import Agent

class Address(BaseModel):
    street: str
    city: str
    country: str
    postal_code: str

class Person(BaseModel):
    name: str
    age: int
    address: Address  # Nested model
    phone_numbers: List[str]  # List of strings
    metadata: Optional[dict] = None

# Configure for nested validation
agent = Agent(
    model="gpt-4",
    response_model=Person,
    # Nested validation settings
    nested_validation={
        # Validate recursively
        "recursive": True,
        # Simplify schema for LLM
        "simplify_schema": True,
        # Provide examples for nested structures
        "include_nested_examples": True,
        # Validation order
        "validate_order": "bottom_up",  # Validate nested first
        # Error handling
        "on_nested_error": "partial",  # Accept partial valid data
        "required_nested_fields": "all"  # All nested fields required
    },
    # Schema generation
    json_schema_mode="compact",  # More compact schema
    schema_depth="full"  # Include full nested structure
)

# PydanticAI now:
# 1. Generates clearer nested schema
# 2. Validates from inside out
# 3. Provides specific error location
# 4. Can accept partial valid data

# Intelligent retry configuration
from pydantic_ai import Agent, RetryStrategy

agent = Agent(
    model="gpt-4",
    response_model=Person,
    # Retry strategy
    retry_strategy=RetryStrategy(
        max_retries=3,
        # Include error in retry prompt
        include_error_details=True,
        error_message_format="detailed",  # Explain what went wrong
        # Exponential backoff
        backoff_strategy="exponential",
        initial_delay=1.0,
        backoff_multiplier=2.0,
        max_delay=10.0,
        # Early stopping
        stop_on_repeated_error=True,
        repeated_error_threshold=2,  # Stop if same error twice
        # Adaptive retry
        adapt_on_error=True,
        lower_temperature_on_retry=True,
        temperature_decay=0.8  # Reduce temp by 20% each retry
    ),
    # Error feedback
    error_feedback={
        "show_field_path": True,  # Show which field failed
        "show_expected_type": True,  # Show what type was expected
        "show_actual_value": True,  # Show what was received
        "suggest_fix": True  # Suggest how to fix
    }
)

# Retry flow:
# Attempt 1: LLM returns invalid data
# → Validation fails with specific error
# → Retry prompt includes error + suggestion
# Attempt 2: LLM sees error, tries to fix
# → If same error, stop early
# → If new error, continue
# → Temperature lowered for more deterministic output