Extract structured data with a schema

Approach

Define the result as a Pydantic model and use client.messages.parse(..., output_format=Model), which returns a validated instance in parsed_output. Keep business validation (allowed values, ranges) in the model, and handle the cases where the input really doesn't contain the data.

Solution

from typing import Literal, Optional

import anthropic
from pydantic import BaseModel, Field

client = anthropic.Anthropic()


class Lead(BaseModel):
    name: str
    email: Optional[str] = None
    company: Optional[str] = None
    plan: Literal["starter", "team", "enterprise", "unknown"]
    seats: Optional[int] = Field(default=None, ge=1, le=100_000)
    wants_demo: bool


def extract_lead(email_text: str) -> Lead:
    response = client.messages.parse(
        model="claude-sonnet-5-5",
        max_tokens=1024,
        system="Extract the sales lead. Use null for anything the email does not state. Never guess emails.",
        messages=[{"role": "user", "content": f"<email>\n{email_text}\n</email>"}],
        output_format=Lead,
    )
    return response.parsed_output


lead = extract_lead("Hi, I'm Priya from Northwind. We'd like Enterprise for about 250 people. Can we see a demo?")
print(lead.model_dump())

Pitfalls

• Making every field required, which pushes the model to invent values; use optional fields and an unknown choice.
• Trusting schema-valid output blindly: check business rules (for example, that the email domain matches the company) in code.
• Putting the raw document in the system prompt; keep untrusted text in the user turn, inside tags.

Variations

• Use output_config={"format": {"type": "json_schema", "schema": {...}}} with messages.create when you don't use Pydantic.
• For many documents, send the requests through the Batch API at a lower price.

Further reading: Structured outputs