Skip to main content

Azure

This notebook goes over how to connect to an Azure hosted OpenAI endpoint

from langchain.chat_models import AzureChatOpenAI
from langchain.schema import HumanMessage
BASE_URL = "https://${TODO}.openai.azure.com"
API_KEY = "..."
DEPLOYMENT_NAME = "chat"
model = AzureChatOpenAI(
    openai_api_base=BASE_URL,
    openai_api_version="2023-05-15",
    deployment_name=DEPLOYMENT_NAME,
    openai_api_key=API_KEY,
    openai_api_type="azure",
)
model(
    [
        HumanMessage(
            content="Translate this sentence from English to French. I love programming."
        )
    ]
)
    AIMessage(content="\n\nJ'aime programmer.", additional_kwargs={})

Model Version​

Azure OpenAI responses contain model property, which is name of the model used to generate the response. However unlike native OpenAI responses, it does not contain the version of the model, which is set on the deplyoment in Azure. This makes it tricky to know which version of the model was used to generate the response, which as result can lead to e.g. wrong total cost calculation with OpenAICallbackHandler.

To solve this problem, you can pass model_version parameter to AzureChatOpenAI class, which will be added to the model name in the llm output. This way you can easily distinguish between different versions of the model.

from langchain.callbacks import get_openai_callback

API Reference:

BASE_URL = "https://{endpoint}.openai.azure.com"
API_KEY = "..."
DEPLOYMENT_NAME = "gpt-35-turbo" # in Azure, this deployment has version 0613 - input and output tokens are counted separately
model = AzureChatOpenAI(
    openai_api_base=BASE_URL,
    openai_api_version="2023-05-15",
    deployment_name=DEPLOYMENT_NAME,
    openai_api_key=API_KEY,
    openai_api_type="azure",
)
with get_openai_callback() as cb:
    model(
        [
            HumanMessage(
                content="Translate this sentence from English to French. I love programming."
            )
        ]
    )
    print(f"Total Cost (USD): ${format(cb.total_cost, '.6f')}") # without specifying the model version, flat-rate 0.002 USD per 1k input and output tokens is used
    Total Cost (USD): $0.000054

We can provide the model version to AzureChatOpenAI constructor. It will get appended to the model name returned by Azure OpenAI and cost will be counted correctly.

model0613 = AzureChatOpenAI(
    openai_api_base=BASE_URL,
    openai_api_version="2023-05-15",
    deployment_name=DEPLOYMENT_NAME,
    openai_api_key=API_KEY,
    openai_api_type="azure",
    model_version="0613"
)
with get_openai_callback() as cb:
    model0613(
        [
            HumanMessage(
                content="Translate this sentence from English to French. I love programming."
            )
        ]
    )
    print(f"Total Cost (USD): ${format(cb.total_cost, '.6f')}")
    Total Cost (USD): $0.000044