The OpenAI Python library provides convenient access to the OpenAI API from applications written in the Python language. It includes a pre-defined set of classes for API resources that initialize themselves dynamically from API responses which makes it compatible with a wide range of versions of the OpenAI API.

You can find usage examples for the OpenAI Python library in our API reference and the OpenAI Cookbook.

You don't need this source code unless you want to modify the package. If you just want to use the package, just run:

pip install --upgrade openai

Install from source with:

python setup.py install

Install dependencies for openai.embeddings_utils:

pip install openai[embeddings]

Install support for Weights & Biases:

pip install openai[wandb] 

Data libraries like numpy and pandas are not installed by default due to their size. They’re needed for some functionality of this library, but generally not for talking to the API. If you encounter a MissingDependencyError, install them with:

pip install openai[datalib]

The library needs to be configured with your account's secret key which is available on the website. Either set it as the OPENAI_API_KEY environment variable before using the library:

export OPENAI_API_KEY='sk-...'

Or set openai.api_key to its value:

importopenaiopenai.api_key="sk-..."# list modelsmodels=openai.Model.list() # print the first model's idprint(models.data[0].id) # create a completioncompletion=openai.Completion.create(model="ada", prompt="Hello world") # print the completionprint(completion.choices[0].text)

All endpoints have a .create method that supports a request_timeout param. This param takes a Union[float, Tuple[float, float]] and will raise an openai.error.Timeout error if the request exceeds that time in seconds (See: https://requests.readthedocs.io/en/latest/user/quickstart/#timeouts).

In order to use the library with Microsoft Azure endpoints, you need to set the api_type, api_base and api_version in addition to the api_key. The api_type must be set to 'azure' and the others correspond to the properties of your endpoint. In addition, the deployment name must be passed as the engine parameter.

importopenaiopenai.api_type="azure"openai.api_key="..."openai.api_base="https://example-endpoint.openai.azure.com"openai.api_version="2023-03-15-preview"# create a completioncompletion=openai.Completion.create(deployment_id="deployment-name", prompt="Hello world") # print the completionprint(completion.choices[0].text)

Please note that for the moment, the Microsoft Azure endpoints can only be used for completion, embedding, and fine-tuning operations. For a detailed example of how to use fine-tuning and other operations using Azure endpoints, please check out the following Jupyter notebooks:

• Using Azure completions
• Using Azure fine-tuning
• Using Azure embeddings

In order to use Microsoft Active Directory to authenticate to your Azure endpoint, you need to set the api_type to "azure_ad" and pass the acquired credential token to api_key. The rest of the parameters need to be set as specified in the previous section.

fromazure.identityimportDefaultAzureCredentialimportopenai# Request credentialdefault_credential=DefaultAzureCredential() token=default_credential.get_token("https://cognitiveservices.azure.com/.default") # Setup parametersopenai.api_type="azure_ad"openai.api_key=token.tokenopenai.api_base="https://example-endpoint.openai.azure.com/"openai.api_version="2023-03-15-preview"# ...

This library additionally provides an openai command-line utility which makes it easy to interact with the API from your terminal. Run openai api -h for usage.

# list models openai api models.list # create a completion openai api completions.create -m ada -p "Hello world"# create a chat completion openai api chat_completions.create -m gpt-3.5-turbo -g user "Hello world"# generate images via DALL·E API openai api image.create -p "two dogs playing chess, cartoon" -n 1

Examples of how to use this Python library to accomplish various tasks can be found in the OpenAI Cookbook. It contains code examples for:

• Classification using fine-tuning
• Clustering
• Code search
• Customizing embeddings
• Question answering from a corpus of documents
• Recommendations
• Visualization of embeddings
• And more

Prior to July 2022, this OpenAI Python library hosted code examples in its examples folder, but since then all examples have been migrated to the OpenAI Cookbook.

Conversational models such as gpt-3.5-turbo can be called using the chat completions endpoint.

importopenaiopenai.api_key="sk-..."# supply your API key however you choosecompletion=openai.ChatCompletion.create(model="gpt-3.5-turbo", messages=[{"role": "user", "content": "Hello world!"}]) print(completion.choices[0].message.content)

In the OpenAI Python library, an embedding represents a text string as a fixed-length vector of floating point numbers. Embeddings are designed to measure the similarity or relevance between text strings.

To get an embedding for a text string, you can use the embeddings method as follows in Python:

importopenaiopenai.api_key="sk-..."# supply your API key however you choose# choose text to embedtext_string="sample text"# choose an embeddingmodel_id="text-similarity-davinci-001"# compute the embedding of the textembedding=openai.Embedding.create(input=text_string, model=model_id)['data'][0]['embedding']

An example of how to call the embeddings method is shown in this get embeddings notebook.

Examples of how to use embeddings are shared in the following Jupyter notebooks:

• Classification using embeddings
• Clustering using embeddings
• Code search using embeddings
• Semantic text search using embeddings
• User and product embeddings
• Zero-shot classification using embeddings
• Recommendation using embeddings

For more information on embeddings and the types of embeddings OpenAI offers, read the embeddings guide in the OpenAI documentation.

Fine-tuning a model on training data can both improve the results (by giving the model more examples to learn from) and reduce the cost/latency of API calls (chiefly through reducing the need to include training examples in prompts).

Examples of fine-tuning are shared in the following Jupyter notebooks:

• Classification with fine-tuning (a simple notebook that shows the steps required for fine-tuning)
• Fine-tuning a model that answers questions about the 2020 Olympics
  • Step 1: Collecting data
  • Step 2: Creating a synthetic Q&A dataset
  • Step 3: Train a fine-tuning model specialized for Q&A

Sync your fine-tunes to Weights & Biases to track experiments, models, and datasets in your central dashboard with:

openai wandb sync

For more information on fine-tuning, read the fine-tuning guide in the OpenAI documentation.

OpenAI provides a Moderation endpoint that can be used to check whether content complies with the OpenAI content policy

importopenaiopenai.api_key="sk-..."# supply your API key however you choosemoderation_resp=openai.Moderation.create(input="Here is some perfectly innocuous text that follows all OpenAI content policies.")

See the moderation guide for more details.

importopenaiopenai.api_key="sk-..."# supply your API key however you chooseimage_resp=openai.Image.create(prompt="two dogs playing chess, oil painting", n=4, size="512x512")

importopenaiopenai.api_key="sk-..."# supply your API key however you choosef=open("path/to/file.mp3", "rb") transcript=openai.Audio.transcribe("whisper-1", f)

Async support is available in the API by prepending a to a network-bound method:

importopenaiopenai.api_key="sk-..."# supply your API key however you chooseasyncdefcreate_completion(): completion_resp=awaitopenai.Completion.acreate(prompt="This is a test", model="davinci")

To make async requests more efficient, you can pass in your own aiohttp.ClientSession, but you must manually close the client session at the end of your program/event loop:

importopenaifromaiohttpimportClientSessionopenai.aiosession.set(ClientSession()) # At the end of your program, close the http sessionawaitopenai.aiosession.get().close()

See the usage guide for more details.

• Python 3.7.1+

In general, we want to support the versions of Python that our customers are using. If you run into problems with any version issues, please let us know on our support page.

This library is forked from the Stripe Python Library.