Skip to main content
Structured outputs ensure your application receives well-defined, machine-readable responses every time. By enforcing formats such as JSON, CSV, or custom Pydantic models, you gain:
BenefitDescription
Type SafetyGuarantees fields are present and correctly typed
Explicit RefusalsModel clearly indicates when it can’t comply
Simpler PromptingRemove guesswork—just specify the format
Exact FormatsEnforce JSON, CSV, or any structure your application requires

​
1. Structured Outputs with Pydantic

Define a Pydantic model for your desired schema and instruct the API to parse directly into that model. 
from pydantic import BaseModel
from openai import OpenAI

client = OpenAI()

class CalendarEvent(BaseModel):
    name: str
    date: str
    participants: list[str]

completion = client.beta.chat.completions.parse(
    model="gpt-4-2024-08-06",
    messages=[
        {"role": "system", "content": "Extract the event information."},
        {"role": "user",   "content": "Alice and Bob are going to a science fair."}
    ],
    response_format=CalendarEvent,
)

event = completion.choices[0].message.parsed
print(event)
# CalendarEvent(name='Science Fair Visit', date='2024-08-10', participants=['Alice', 'Bob'])
Make sure you’re using the openai Python package version that supports .parse() (e.g., v0.27+).
See Pydantic documentation for detailed model usage.

​
2. Step-by-Step Reasoning as Structured Output

You can capture a chain of thought or intermediate steps by defining a Pydantic model that holds your reasoning. 
from pydantic import BaseModel
from openai import OpenAI

client = OpenAI()

class MathReasoning(BaseModel):
    steps: list[str]
    final_answer: str

completion = client.beta.chat.completions.parse(
    model="gpt-4-2024-08-06",
    messages=[
        {"role": "system", "content": "You are a helpful math tutor. Guide the user."},
        {"role": "user",   "content": "How can I solve 8x + 7 = -23?"}
    ],
    response_format=MathReasoning,
)

math = completion.choices[0].message.parsed
print("Steps:", math.steps)
print("Answer:", math.final_answer)

​
3. Building a Recipe Generator with JSON Output

Prompt the model to output strict JSON so downstream services can ingest it without additional parsing. 
from openai import OpenAI

client = OpenAI(api_key="sk-REPLACE_WITH_YOUR_KEY")

def recipe_gen(ingredients: list[str]) -> str:
    messages = [{"role": "user", "content": ing} for ing in ingredients]
    messages.extend([
        {"role": "system",    "content": "JSON Format Required"},
        {"role": "assistant", "content": "You are a high-end chef. Generate a recipe in JSON!"}
    ])

    response = client.chat.completions.create(
        model="gpt-4",
        messages=messages,
        max_tokens=300,
        temperature=0.9
    )
    return response.choices[0].message.content

if __name__ == "__main__":
    ingredients = []
    while True:
        ing = input("Enter an ingredient (or type done): ")
        if ing.strip().lower() == "done":
            break
        ingredients.append(ing)

    print(recipe_gen(ingredients))
Run the script: 
$ python recipe_generator.py
Enter an ingredient (or type done): chocolate
Enter an ingredient (or type done): grapes
Enter an ingredient (or type done): pizza dough
Enter an ingredient (or type done): done
Example JSON response: 
{
  "recipe": {
    "title": "Chocolate-Infused Grape Pizza Tacos",
    "servings": 4,
    "ingredients": [
      { "name": "Corn tortillas",   "quantity": 8,   "unit": "pieces" },
      { "name": "Pizza dough",      "quantity": 1,   "unit": "pound" },
      { "name": "Dark chocolate",   "quantity": 100, "unit": "grams", "type": "finely chopped" },
      { "name": "Red grapes",       "quantity": 200, "unit": "grams", "type": "halved" },
      { "name": "Mozzarella cheese","quantity": 200, "unit": "grams", "type": "shredded" },
      { "name": "Ricotta cheese",   "quantity": 100, "unit": "grams" },
      { "name": "Fresh basil leaves","quantity": 12,  "unit": "leaves" }
    ]
  }
}
Never commit your API keys to public repositories. Use environment variables or a secure vault.

​
4. Switching to CSV Output

Simply update the system prompt to request CSV, then process the comma-separated response. 
from openai import OpenAI

client = OpenAI(api_key="sk-REPLACE_WITH_YOUR_KEY")

def recipe_gen_csv(ingredients: list[str]) -> str:
    messages = [{"role": "user", "content": ing} for ing in ingredients]
    messages.extend([
        {"role": "system",    "content": "CSV Format Required"},
        {"role": "assistant", "content": "You are a high-end chef. Generate a CSV recipe!"}
    ])

    response = client.chat.completions.create(
        model="gpt-4",
        messages=messages,
        max_tokens=300,
        temperature=0.9
    )
    return response.choices[0].message.content

if __name__ == "__main__":
    ingredients = []
    while True:
        ing = input("Enter an ingredient (or type done): ")
        if ing.strip().lower() == "done":
            break
        ingredients.append(ing)

    print(recipe_gen_csv(ingredients))
Sample CSV output: 
Ingredient,Quantity,Preparation
Chicken Breast,2 pieces,Cut into cubes
Potatoes,2 medium,Peeled and diced
Eggs,4,Hard-boiled and sliced
Olive Oil,2 Tbsp,For cooking
Garlic,2 cloves,Minced
Onion,1 medium,Finely chopped
Salt,1 tsp,To taste
Black Pepper,1/2 tsp,To taste
Paprika,1/2 tsp,To taste
Fresh Parsley,2 tbsp,Chopped for garnish
You can even mix formats—return cooking steps as CSV: 
Step 1,Heat olive oil in a large pan over medium heat.
Step 2,Add chicken and onion; sauté until translucent.
Step 3,Add garlic, potatoes, and paprika; cook until tender.
Step 4,Stir in eggs until heated through.
Step 5,Garnish with parsley and serve.
By specifying your desired output format in the system and assistant messages, you guarantee consistency and simplify downstream processing. For more advanced examples, see the OpenAI Structured Outputs Guide.

Watch Video