Structured Outputs, JSON Schema & Pydantic

1. The Evolution of Structured Generation

2. How Constrained Decoding Works Under the Hood

When strict: true is enabled, the inference engine (such as OpenAI's infrastructure, Outlines, or SGLang) converts your JSON Schema into a Deterministic Finite Automaton (DFA):

At each sampling step, logits for all invalid tokens that would violate the grammar are set to negative infinity (-inf). It is literally impossible for the model to emit a trailing comma, missing bracket, or unexpected string!

3. Production Pydantic Extraction with OpenAI

By leveraging OpenAI's client.beta.chat.completions.parse() method, you can pass native Pydantic models and receive fully typed, validated Python objects:

from typing import List, Optional
from pydantic import BaseModel, Field
from openai import OpenAI

client = OpenAI()

# 1. Define Strict Pydantic Schema
class InvoiceItem(BaseModel):
    description: str = Field(description="Name or description of the product/service")
    quantity: int = Field(description="Number of units purchased", ge=1)
    unit_price: float = Field(description="Price per single unit in USD")
    total: float = Field(description="Line item subtotal (quantity * unit_price)")

class InvoiceExtraction(BaseModel):
    invoice_number: str
    vendor_name: str
    date: str
    items: List[InvoiceItem]
    tax_amount: Optional[float] = 0.0
    total_amount: float

# 2. Extract from Raw Unstructured Text
raw_email = """
Thanks for your business! Invoice #INV-2026-9812
From: CloudTech Solutions LLC
Date: October 14, 2026
Items:
- Enterprise Kubernetes Cluster (qty: 2) at $1,250.00 each -> $2,500.00
- Dedicated Support SLA (qty: 1) at $500.00 -> $500.00
Tax: $150.00
Total Due: $3,150.00
"""

completion = client.beta.chat.completions.parse(
    model="gpt-4o-2024-08-06",
    messages=[
        {"role": "system", "content": "Extract invoice details into the structured format."},
        {"role": "user", "content": raw_email}
    ],
    response_format=InvoiceExtraction
)

# 3. Access strongly-typed properties with full IDE autocomplete!
invoice_data: InvoiceExtraction = completion.choices[0].message.parsed
print(f"Vendor: {invoice_data.vendor_name}")
print(f"Total Amount: ${invoice_data.total_amount:,.2f}")
for item in invoice_data.items:
    print(f"  - {item.description}: {item.quantity} x ${item.unit_price} = ${item.total}")

4. Best Practices for Strict Schemas