Pydantic — Data Validation for AI Agent Pipelines

What it is

Pydantic is Python's most widely used data validation library. It uses Python type hints to define data schemas and validates input automatically at runtime. In the AI ecosystem, Pydantic is the foundation for tool definitions in agent frameworks, structured output parsing from LLMs, and API request/response validation.

The library targets Python developers building AI pipelines, API services, or any application where data integrity matters. It powers the schema layer of FastAPI, LangChain, LlamaIndex, and most major AI agent frameworks.

How it saves time or tokens

Pydantic eliminates manual validation code. Instead of writing if-else checks for every field, you declare a model class with type annotations and Pydantic handles coercion, constraint checking, and error reporting. This reduces boilerplate significantly.

For AI agents, Pydantic models serve as tool parameter schemas. The agent framework generates the JSON schema from the model, sends it to the LLM, and validates the LLM's response against the same model. This catches malformed outputs before they reach your application logic.

How to use

1. Install with pip install pydantic.
2. Define a model class inheriting from BaseModel with typed fields and optional validators.
3. Instantiate the model with data -- Pydantic validates automatically and raises ValidationError on invalid input.

Example

from pydantic import BaseModel, Field
from typing import Optional

class UserProfile(BaseModel):
    name: str = Field(description='Full name')
    age: int = Field(ge=0, le=150, description='Age in years')
    email: str = Field(pattern=r'^[\w.-]+@[\w.-]+\.\w+$')
    bio: Optional[str] = None

# Validates automatically
user = UserProfile(name='Alice', age=30, email='alice@example.com')
print(user.model_dump_json())

# Raises ValidationError
try:
    bad = UserProfile(name='Bob', age=-5, email='not-email')
except Exception as e:
    print(e)

Related on TokRepo

• AI tools for coding -- Python libraries for AI development
• AI tools for agents -- Agent frameworks that rely on Pydantic

Common pitfalls

• Pydantic v2 is a major rewrite with breaking changes from v1. Methods like .dict() are renamed to .model_dump(). Check your Pydantic version when following tutorials.
• Pydantic coerces types by default. A string '42' becomes integer 42 silently. Use strict=True on the model config if you need exact type matching.
• Nested models with circular references require model_rebuild() after all models are defined. Forgetting this causes PydanticUndefined errors.