Structured Output
Get typed Python objects from LLMs using Pydantic, with automatic schema enforcement and retries across providers.
Core idea
Pass response_model=YourPydanticModel to generate() and receive a validated instance (or a validation error after retries).
Basic example
from pydantic import BaseModel
from abstractcore import create_llm
class Person(BaseModel):
name: str
age: int
email: str
llm = create_llm("openai", model="gpt-5-mini")
person = llm.generate(
"Extract: John Doe, 35 years old, john@example.com",
response_model=Person,
)
assert isinstance(person, Person)
print(person.name, person.age, person.email)Note: Structured output is designed for non-streaming calls (use stream=False, the default).
Automatic validation + retry
When a provider cannot enforce the schema natively, AbstractCore can retry with validation feedback.
from pydantic import BaseModel
from abstractcore import create_llm
from abstractcore.structured import FeedbackRetry
class Company(BaseModel):
name: str
website: str
llm = create_llm("ollama", model="qwen3:4b-instruct")
company = llm.generate(
"Extract: ACME Corp, https://acme.example",
response_model=Company,
retry_strategy=FeedbackRetry(max_attempts=3),
)Provider notes (native vs prompted)
AbstractCore selects a strategy automatically based on provider/model capabilities.
| Provider | Typical strategy | How |
|---|---|---|
| OpenAI | Native | Schema passed to API (response_format) |
| Anthropic | Native | Schema enforced via tool-calling mechanism |
| Ollama | Native | Schema passed via format |
| LMStudio / OpenAI-compatible | Native (when supported by backend) | OpenAI-style schema payload |
| MLX / HuggingFace | Prompted + retry | Schema prompt + validation feedback loop (or native when available) |
Tools + structured output (hybrid)
If you pass both tools=[...] and response_model=..., AbstractCore can use a hybrid flow (tool use, then structured follow-up).
For complex agent workflows, keep tool execution host-owned (execute_tools=false, default) and validate schemas at the boundary.