LangChain Module 3: Structured Output with Pydantic

Step 1: Define the schema

Use a Pydantic model with typed fields. Field descriptions are sent to the model, so write them as instructions that explain what each field means.

from pydantic import BaseModel, Field

class Movie(BaseModel):
    """Details about a movie."""
    title: str = Field(description="The movie's title")
    year: int = Field(description="The year it was released")
    director: str = Field(description="The director's full name")

Step 2: Use with_structured_output

Wrap a model with the schema. The wrapped model returns an instance of your class instead of a message, so you can use attributes directly.

structured_model = model.with_structured_output(Movie)

movie = structured_model.invoke("Tell me about the movie Inception.")
print(movie.title, movie.year, movie.director)

Step 3: Use response_format with agents

When you build an agent, pass the schema as response_format. The validated object is returned under the structured_response key of the result. The tools lesson explains agents in detail.

from langchain.agents import create_agent

agent = create_agent(model=model, response_format=Movie)

result = agent.invoke(
    {"messages": [{"role": "user", "content": "Tell me about the movie Inception."}]}
)
print(result["structured_response"])

Step 4: Know the schema options and limits

Pydantic models return validated instances, while dataclass, TypedDict, and JSON Schema inputs return dictionaries. A JSON Schema dictionary must be wrapped in an explicit strategy such as ToolStrategy or ProviderStrategy, and needs a top-level title and description. Use Pydantic when you can, since it also validates.

Step 5: Validate meaning, not just shape

A schema guarantees the structure, not the truth. Add your own checks, such as plausible ranges and cross-checks with trusted data, and decide what to do with outputs that fail.