Instructor is the most popular Python library for working with structured outputs from large language models (LLMs), boasting over 1 million monthly downloads. Built on top of Pydantic, it provides a simple, transparent, and user-friendly API to manage validation, retries, and streaming responses. Get ready to supercharge your LLM workflows with the community's top choice!
If your company uses Instructor a lot, we'd love to have your logo on our website! Please fill out this form
- Response Models: Specify Pydantic models to define the structure of your LLM outputs
- Retry Management: Easily configure the number of retry attempts for your requests
- Validation: Ensure LLM responses conform to your expectations with Pydantic validation
- Streaming Support: Work with Lists and Partial responses effortlessly
- Flexible Backends: Seamlessly integrate with various LLM providers beyond OpenAI
- Support in many Languages: We support many languages including Python, TypeScript, Ruby, Go, and Elixir
Install Instructor with a single command:
pip install -U instructorNow, let's see Instructor in action with a simple example:
import instructor
from pydantic import BaseModel
from openai import OpenAI
# Define your desired output structure
class UserInfo(BaseModel):
name: str
age: int
# Patch the OpenAI client
client = instructor.from_openai(OpenAI())
# Extract structured data from natural language
user_info = client.chat.completions.create(
model="gpt-4o-mini",
response_model=UserInfo,
messages=[{"role": "user", "content": "John Doe is 30 years old."}],
)
print(user_info.name)
#> John Doe
print(user_info.age)
#> 30Instructor provides a powerful hooks system that allows you to intercept and log various stages of the LLM interaction process. Here's a simple example demonstrating how to use hooks:
import instructor
from openai import OpenAI
from pydantic import BaseModel
class UserInfo(BaseModel):
name: str
age: int
# Initialize the OpenAI client with Instructor
client = instructor.from_openai(OpenAI())
# Define hook functions
def log_kwargs(**kwargs):
print(f"Function called with kwargs: {kwargs}")
def log_exception(exception: Exception):
print(f"An exception occurred: {str(exception)}")
client.on("completion:kwargs", log_kwargs)
client.on("completion:error", log_exception)
user_info = client.chat.completions.create(
model="gpt-4o-mini",
response_model=UserInfo,
messages=[
{"role": "user", "content": "Extract the user name: 'John is 20 years old'"}
],
)
"""
{
'args': (),
'kwargs': {
'messages': [
{
'role': 'user',
'content': "Extract the user name: 'John is 20 years old'",
}
],
'model': 'gpt-4o-mini',
'tools': [
{
'type': 'function',
'function': {
'name': 'UserInfo',
'description': 'Correctly extracted `UserInfo` with all the required parameters with correct types',
'parameters': {
'properties': {
'name': {'title': 'Name', 'type': 'string'},
'age': {'title': 'Age', 'type': 'integer'},
},
'required': ['age', 'name'],
'type': 'object',
},
},
}
],
'tool_choice': {'type': 'function', 'function': {'name': 'UserInfo'}},
},
}
"""
print(f"Name: {user_info.name}, Age: {user_info.age}")
#> Name: John, Age: 20This example demonstrates:
- A pre-execution hook that logs all kwargs passed to the function.
- An exception hook that logs any exceptions that occur during execution.
The hooks provide valuable insights into the function's inputs and any errors, enhancing debugging and monitoring capabilities.
import instructor
from anthropic import Anthropic
from pydantic import BaseModel
class User(BaseModel):
name: str
age: int
client = instructor.from_anthropic(Anthropic())
# note that client.chat.completions.create will also work
resp = client.messages.create(
model="claude-3-opus-20240229",
max_tokens=1024,
system="You are a world class AI that excels at extracting user data from a sentence",
messages=[
{
"role": "user",
"content": "Extract Jason is 25 years old.",
}
],
response_model=User,
)
assert isinstance(resp, User)
assert resp.name == "Jason"
assert resp.age == 25Make sure to install cohere and set your system environment variable with export CO_API_KEY=<YOUR_COHERE_API_KEY>.
pip install cohere
import instructor
import cohere
from pydantic import BaseModel
class User(BaseModel):
name: str
age: int
client = instructor.from_cohere(cohere.Client())
# note that client.chat.completions.create will also work
resp = client.chat.completions.create(
model="command-r-plus",
max_tokens=1024,
messages=[
{
"role": "user",
"content": "Extract Jason is 25 years old.",
}
],
response_model=User,
)
assert isinstance(resp, User)
assert resp.name == "Jason"
assert resp.age == 25Make sure you install the Google AI Python SDK. You should set a GOOGLE_API_KEY environment variable with your API key.
Gemini tool calling also requires jsonref to be installed.
pip install google-generativeai jsonref
import instructor
import google.generativeai as genai
from pydantic import BaseModel
class User(BaseModel):
name: str
age: int
# genai.configure(api_key=os.environ["API_KEY"]) # alternative API key configuration
client = instructor.from_gemini(
client=genai.GenerativeModel(
model_name="models/gemini-1.5-flash-latest", # model defaults to "gemini-pro"
),
mode=instructor.Mode.GEMINI_JSON,
)Alternatively, you can call Gemini from the OpenAI client. You'll have to setup gcloud, get setup on Vertex AI, and install the Google Auth library.
pip install google-authimport google.auth
import google.auth.transport.requests
import instructor
from openai import OpenAI
from pydantic import BaseModel
creds, project = google.auth.default()
auth_req = google.auth.transport.requests.Request()
creds.refresh(auth_req)
# Pass the Vertex endpoint and authentication to the OpenAI SDK
PROJECT = 'PROJECT_ID'
LOCATION = (
'LOCATION' # https://cloud.google.com/vertex-ai/generative-ai/docs/learn/locations
)
base_url = f'https://{LOCATION}-aiplatform.googleapis.com/v1beta1/projects/{PROJECT}/locations/{LOCATION}/endpoints/openapi'
client = instructor.from_openai(
OpenAI(base_url=base_url, api_key=creds.token), mode=instructor.Mode.JSON
)
# JSON mode is req'd
class User(BaseModel):
name: str
age: int
resp = client.chat.completions.create(
model="google/gemini-1.5-flash-001",
max_tokens=1024,
messages=[
{
"role": "user",
"content": "Extract Jason is 25 years old.",
}
],
response_model=User,
)
assert isinstance(resp, User)
assert resp.name == "Jason"
assert resp.age == 25import instructor
from litellm import completion
from pydantic import BaseModel
class User(BaseModel):
name: str
age: int
client = instructor.from_litellm(completion)
resp = client.chat.completions.create(
model="claude-3-opus-20240229",
max_tokens=1024,
messages=[
{
"role": "user",
"content": "Extract Jason is 25 years old.",
}
],
response_model=User,
)
assert isinstance(resp, User)
assert resp.name == "Jason"
assert resp.age == 25This was the dream of Instructor but due to the patching of OpenAI, it wasn't possible for me to get typing to work well. Now, with the new client, we can get typing to work well! We've also added a few create_* methods to make it easier to create iterables and partials, and to access the original completion.
import openai
import instructor
from pydantic import BaseModel
class User(BaseModel):
name: str
age: int
client = instructor.from_openai(openai.OpenAI())
user = client.chat.completions.create(
model="gpt-4-turbo-preview",
messages=[
{"role": "user", "content": "Create a user"},
],
response_model=User,
)Now if you use an IDE, you can see the type is correctly inferred.
This will also work correctly with asynchronous clients.
import openai
import instructor
from pydantic import BaseModel
client = instructor.from_openai(openai.AsyncOpenAI())
class User(BaseModel):
name: str
age: int
async def extract():
return await client.chat.completions.create(
model="gpt-4-turbo-preview",
messages=[
{"role": "user", "content": "Create a user"},
],
response_model=User,
)Notice that simply because we return the create method, the extract() function will return the correct user type.
You can also return the original completion object
import openai
import instructor
from pydantic import BaseModel
client = instructor.from_openai(openai.OpenAI())
class User(BaseModel):
name: str
age: int
user, completion = client.chat.completions.create_with_completion(
model="gpt-4-turbo-preview",
messages=[
{"role": "user", "content": "Create a user"},
],
response_model=User,
)
In order to handle streams, we still support Iterable[T] and Partial[T] but to simplify the type inference, we've added create_iterable and create_partial methods as well!
import openai
import instructor
from pydantic import BaseModel
client = instructor.from_openai(openai.OpenAI())
class User(BaseModel):
name: str
age: int
user_stream = client.chat.completions.create_partial(
model="gpt-4-turbo-preview",
messages=[
{"role": "user", "content": "Create a user"},
],
response_model=User,
)
for user in user_stream:
print(user)
#> name=None age=None
#> name=None age=None
#> name=None age=None
#> name=None age=None
#> name=None age=None
#> name=None age=None
#> name='John Doe' age=None
#> name='John Doe' age=None
#> name='John Doe' age=None
#> name='John Doe' age=30
#> name='John Doe' age=30
# name=None age=None
# name='' age=None
# name='John' age=None
# name='John Doe' age=None
# name='John Doe' age=30Notice now that the type inferred is Generator[User, None]
We get an iterable of objects when we want to extract multiple objects.
import openai
import instructor
from pydantic import BaseModel
client = instructor.from_openai(openai.OpenAI())
class User(BaseModel):
name: str
age: int
users = client.chat.completions.create_iterable(
model="gpt-4-turbo-preview",
messages=[
{"role": "user", "content": "Create 2 users"},
],
response_model=User,
)
for user in users:
print(user)
#> name='John Doe' age=30
#> name='Jane Doe' age=28
# User(name='John Doe', age=30)
# User(name='Jane Smith', age=25)
We invite you to contribute to evals in pytest as a way to monitor the quality of the OpenAI models and the instructor library. To get started check out the evals for Anthropic and OpenAI and contribute your own evals in the form of pytest tests. These evals will be run once a week and the results will be posted.
If you want to help, checkout some of the issues marked as good-first-issue or help-wanted found here. They could be anything from code improvements, a guest blog post, or a new cookbook.
Here's a quick list of commands that you can run to get started. We're using uv to manage our dependencies so make sure you have that installed.
-
uv sync --all-extras --group <dependency groups you'd like to install>: This should install all the dependencies for the project usinguv, make sure to install the specific dependencies that you'd like to install -
uv run pytest: This runs the tests inpytest. If you're pushing up a new PR, make sure that you've written some tests and that they're passing locally for you
We use ruff and pyright for linting and type checking so make sure those are passing when you push up a PR. You can check pyright by running uv run pyright and ruff with uv run ruff check locally.
We also provide some added CLI functionality for easy convenience:
-
instructor jobs: This helps with the creation of fine-tuning jobs with OpenAI. Simple useinstructor jobs create-from-file --helpto get started creating your first fine-tuned GPT-3.5 model -
instructor files: Manage your uploaded files with ease. You'll be able to create, delete and upload files all from the command line -
instructor usage: Instead of heading to the OpenAI site each time, you can monitor your usage from the CLI and filter by date and time period. Note that usage often takes ~5-10 minutes to update from OpenAI's side
This project is licensed under the terms of the MIT License.
![@ellipsis-dev[bot]](https://avatars.githubusercontent.com/in/64358?s=64&v=4){kind=small}
