Pydantic + LLM Structured Outputs Cheat Sheet

What is Pydantic?

A Python library for data validation and serialization that automatically:

• Validates data types
• Converts types when possible
• Provides clear error messages
• Serializes to/from JSON, dictionaries

Basic Pydantic Model

from pydantic import BaseModel

class Evaluation(BaseModel):
    is_acceptable: bool
    feedback: str

Using with LLMs

The Problem

• LLMs return unstructured text
• Hard to parse and validate
• Inconsistent formats

The Solution

Use Pydantic models to enforce structure!

response = gemini.beta.chat.completions.parse(
    model="gemini-2.0-flash", 
    messages=messages, 
    response_format=Evaluation  # Your Pydantic model
)

# Returns validated Evaluation object
print(response.is_acceptable)  # bool
print(response.feedback)       # str

How It Works Under the Hood

1. Your Pydantic Model → JSON Schema (automatic conversion)
2. JSON Schema → LLM API (as instructions)
3. LLM generates → JSON matching schema
4. Library parses → Validated Pydantic object

Example JSON Schema Generated

{
  "type": "object",
  "properties": {
    "is_acceptable": {"type": "boolean"},
    "feedback": {"type": "string"}
  },
  "required": ["is_acceptable", "feedback"]
}

Benefits

✅ Guaranteed structure - Always get expected fields
✅ Type safety - Fields are correct types
✅ Error handling - Validation catches format errors
✅ Easy processing - Use data immediately in code
✅ No manual parsing - Library handles conversion

Popular Libraries

• Instructor - Easy structured outputs with OpenAI/Anthropic
• Marvin - Another structured LLM library
• LangChain - Built-in Pydantic output parsers
• Gemini - Native response_format support

Quick Example

from pydantic import BaseModel

class MovieReview(BaseModel):
    title: str
    rating: int  # 1-10
    summary: str
    recommended: bool

# LLM automatically returns structured data
review = api_call(response_format=MovieReview)
print(f"Rating: {review.rating}")  # Guaranteed integer

Key Takeaway

Pydantic models become “contracts” - you define what you want, and the LLM API ensures you get exactly that structure. No more messy text parsing!