Structured Output with Pydantic and LLMs

For Python applications calling LLMs for structured output, Pydantic is the right schema layer. Type-safe, validates server responses, plays well with FastAPI / agentic frameworks. Combined with vLLM's guided decoding, you get end-to-end type-safe LLM output.

Pattern

from pydantic import BaseModel, Field
from openai import OpenAI

class Invoice(BaseModel):
    vendor: str
    invoice_date: str = Field(..., description="ISO date")
    total: float
    currency: str = Field(..., pattern="^(GBP|USD|EUR)$")
    line_items: list[dict] = Field(default_factory=list)

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")
resp = client.beta.chat.completions.parse(
    model="meta-llama/Meta-Llama-3.1-8B-Instruct",
    messages=[
        {"role": "system", "content": "Extract invoice data."},
        {"role": "user", "content": invoice_text},
    ],
    response_format=Invoice,  # auto-generates JSON schema
)
invoice: Invoice = resp.choices[0].message.parsed

Setup

• Pydantic v2 (use BaseModel; v1 is legacy)
• OpenAI Python SDK 1.40+ supports response_format=BaseModel directly
• vLLM 0.6+ accepts JSON schemas via response_format
• For older OpenAI SDK: convert via model_json_schema() manually

Validation

Pydantic validates the server response on parse. Validation failures (extremely rare with guided decoding) raise ValidationError — treat as a logic bug, not a fallback path. With guided decoding active, output is valid by construction.

Use Pydantic Field validators for runtime constraints beyond JSON Schema:

• Custom regex patterns
• Numeric bounds
• Enum values
• Cross-field consistency checks

Verdict

For Python production LLM apps with structured outputs, Pydantic + vLLM guided decoding + OpenAI SDK's native Pydantic support is the right stack. Type-safe end-to-end, IDE-friendly, well-tested ecosystem. The standard pattern in 2026 production Python AI apps.