Structured Outputs¶

Pydantic Models¶

The @prompt decorator will respect the return type annotation of the decorated function. This can be any type supported by pydantic including a pydantic model. See the Pydantic docs for more information about models.

from magentic import prompt
from pydantic import BaseModel


class Superhero(BaseModel):
    name: str
    age: int
    power: str
    enemies: list[str]


@prompt("Create a Superhero named {name}.")
def create_superhero(name: str) -> Superhero: ...


create_superhero("Garden Man")
# Superhero(name='Garden Man', age=30, power='Control over plants', enemies=['Pollution Man', 'Concrete Woman'])

Using `Field`¶

With pydantic's BaseModel, you can use Field to provide additional information for individual attributes, such as a description.

from magentic import prompt
from pydantic import BaseModel


class Superhero(BaseModel):
    name: str
    age: int = Field(
        description="The age of the hero, which could be much older than humans."
    )
    power: str = Field(examples=["Runs really fast"])
    enemies: list[str]


@prompt("Create a Superhero named {name}.")
def create_superhero(name: str) -> Superhero: ...

ConfigDict¶

Pydantic also supports configuring the BaseModel by setting the model_config attribute. Magentic extends pydantic's ConfigDict class to add the following additional configuration options

openai_strict: bool Indicates whether to use OpenAI's Structured Outputs feature.

See the pydantic Configuration docs for the inherited configuration options.

from magentic import prompt, ConfigDict
from pydantic import BaseModel


class Superhero(BaseModel):
    model_config = ConfigDict(openai_strict=True)

    name: str
    age: int
    power: str
    enemies: list[str]


@prompt("Create a Superhero named {name}.")
def create_superhero(name: str) -> Superhero: ...


create_superhero("Garden Man")

JSON Schema¶

OpenAI Structured Outputs

Setting openai_strict=True results in a different JSON schema than that from .model_json_schema() being sent to the LLM. Use openai.pydantic_function_tool(Superhero) to generate the JSON schema in this case.

You can generate the JSON schema for the pydantic model using the .model_json_schema() method. This is what is sent to the LLM.

Running Superhero.model_json_schema() for the above definition reveals the following JSON schema

{
    "properties": {
        "name": {"title": "Name", "type": "string"},
        "age": {
            "description": "The age of the hero, which could be much older than humans.",
            "title": "Age",
            "type": "integer",
        },
        "power": {
            "examples": ["Runs really fast"],
            "title": "Power",
            "type": "string",
        },
        "enemies": {"items": {"type": "string"}, "title": "Enemies", "type": "array"},
    },
    "required": ["name", "age", "power", "enemies"],
    "title": "Superhero",
    "type": "object",
}

If a StructuredOutputError is raised often, this indicates that the LLM is failing to match the schema. The traceback for these errors includes the underlying pydantic ValidationError which shows in what way the received response was invalid. To combat these errors there are several options

Add descriptions or examples for individual fields to demonstrate valid values.
Simplify the output schema, including using more flexible types (e.g. str instead of datetime) or allowing fields to be nullable with | None.
Switch to a "more intelligent" LLM. See Configuration for how to do this.

Python Types¶

Regular Python types can also be used as the function return type.

from magentic import prompt
from pydantic import BaseModel, Field


class Superhero(BaseModel):
    name: str
    age: int
    power: str
    enemies: list[str]


garden_man = Superhero(
    name="Garden Man",
    age=30,
    power="Control over plants",
    enemies=["Pollution Man", "Concrete Woman"],
)


@prompt("Return True if {hero.name} will be defeated by enemies {hero.enemies}")
def will_be_defeated(hero: Superhero) -> bool: ...


hero_defeated = will_be_defeated(garden_man)
print(hero_defeated)
# > True

Chain-of-Thought Prompting¶

Using a simple Python type as the return annotation might result in poor results as the LLM has no time to arrange its thoughts before answering. To allow the LLM to work through this "chain of thought" you can instead return a pydantic model with initial fields for explaining the final response.

from magentic import prompt
from pydantic import BaseModel, Field


class ExplainedDefeated(BaseModel):
    explanation: str = Field(
        description="Describe the battle between the hero and their enemy."
    )
    defeated: bool = Field(description="True if the hero was defeated.")


class Superhero(BaseModel):
    name: str
    age: int
    power: str
    enemies: list[str]


@prompt("Return True if {hero.name} will be defeated by enemies {hero.enemies}")
def will_be_defeated(hero: Superhero) -> ExplainedDefeated: ...


garden_man = Superhero(
    name="Garden Man",
    age=30,
    power="Control over plants",
    enemies=["Pollution Man", "Concrete Woman"],
)

hero_defeated = will_be_defeated(garden_man)
print(hero_defeated.defeated)
# > True
print(hero_defeated.explanation)
# > 'Garden Man is an environmental hero who fights against Pollution Man ...'

Explained¶

Using chain-of-thought is a common approach to improve the output of the model, so a generic Explained model might be generally useful. The description or example parameters of Field can be used to demonstrate the desired style and detail of the explanations.

from typing import Generic, TypeVar

from magentic import prompt
from pydantic import BaseModel, Field


T = TypeVar("T")


class Explained(BaseModel, Generic[T]):
    explanation: str = Field(description="Explanation of how the value was determined.")
    value: T


@prompt("Return True if {hero.name} will be defeated by enemies {hero.enemies}")
def will_be_defeated(hero: Superhero) -> Explained[bool]: ...