AI Agents with LangGraph

Structured Output Demo

You’ll use the ChatOpenAI model for this demo, so make sure you have your API key in your .env file in the root of your project:

OPENAI_API_KEY=<your-api-key>

Open the empty structured.ipynb notebook in the Starter project. Then load your API key:

from dotenv import load_dotenv
load_dotenv()

Set up the ChatOpenAI LLM:

import os
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
  api_key=os.getenv("OPENAI_API_KEY"),
  temperature=1.0
)

Since you want a little more randomness in the output, you set the temperature to 1.0. If you find that’s affecting the quality of the structured output, you can tone it down.

Then, define a Pydantic model as the instruction section showed:

from langchain_core.pydantic_v1 import BaseModel, Field

class Person(BaseModel):
  """Profile of a human."""

  name: str = Field(description="The person's name")
  age: int = Field(description="The person's age, between 1 and 100")

Give the Pydantic model to your large language model:

structured_llm = llm.with_structured_output(Person)
structured_llm.invoke("Create a random character for a story")

And there you have a Pydantic model. Rerun it a few times. This doesn’t seem as random as it’s supposed to be. A bit better prompting might improve that, but that’s a challenge for another day.

Next, create a TypedDict version for the same class:

from typing_extensions import Annotated, TypedDict

class Person(TypedDict):
  """Profile of a human."""

  name: Annotated[str, ..., "The person's name"]
  age: Annotated[int, ..., "The person's age"]

Using Annotated lets you add some metadata to help the LLM. Normal comments would have worked, too. The ... means that the field isn’t optional.

Run the LLM as before; this time, you get a dictionary as the output.