Azure OpenAI binding spec

Detailed documentation on the Azure OpenAI binding component

Component format

To setup an Azure OpenAI binding create a component of type bindings.azure.openai. See this guide on how to create and apply a binding configuration. See this for the documentation for Azure OpenAI Service.

apiVersion: dapr.io/v1alpha1
kind: Component
metadata:
  name: <NAME>
spec:
  type: bindings.azure.openai
  version: v1
  metadata:
  - name: apiKey # Required
    value: "1234567890abcdef"
  - name: endpoint # Required
    value: "https://myopenai.openai.azure.com"

Warning

The above example uses apiKey as a plain string. It is recommended to use a secret store for the secrets as described here.

Spec metadata fields

Azure Active Directory (AAD) authentication

The Azure OpenAI binding component supports authentication using all Azure Active Directory mechanisms. For further information and the relevant component metadata fields to provide depending on the choice of AAD authentication mechanism, see the docs for authenticating to Azure.

Example Configuration

apiVersion: dapr.io/v1alpha1
kind: component
metadata:
  name: <NAME>
spec:
  type: bindings.azure.openai
  version: v1
  metadata:
  - name: endpoint
    value: "https://myopenai.openai.azure.com"
  - name: azureTenantId
    value: "***"
  - name: azureClientId
    value: "***"
  - name: azureClientSecret
    value: "***"

Binding support

This component supports output binding with the following operations:

• completion : Completion API
• chat-completion : Chat Completion API
• get-embedding : Embedding API

Completion API

To call the completion API with a prompt, invoke the Azure OpenAI binding with a POST method and the following JSON body:

{
  "operation": "completion",
  "data": {
    "deploymentId": "my-model",
    "prompt": "A dog is",
    "maxTokens":5
    }
}

The data parameters are:

• deploymentId - string that specifies the model deployment ID to use.
• prompt - string that specifies the prompt to generate completions for.
• maxTokens - (optional) defines the max number of tokens to generate. Defaults to 16 for completion API.
• temperature - (optional) defines the sampling temperature between 0 and 2. Higher values like 0.8 make the output more random, while lower values like 0.2 make it more focused and deterministic. Defaults to 1.0 for completion API.
• topP - (optional) defines the sampling temperature. Defaults to 1.0 for completion API.
• n - (optional) defines the number of completions to generate. Defaults to 1 for completion API.
• presencePenalty - (optional) Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far, increasing the model’s likelihood to talk about new topics. Defaults to 0.0 for completion API.
• frequencyPenalty - (optional) Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model’s likelihood to repeat the same line verbatim. Defaults to 0.0 for completion API.

Read more about the importance and usage of these parameters in the Azure OpenAI API documentation.

Examples

• Linux

curl -d '{ "data": {"deploymentId: "my-model" , "prompt": "A dog is ", "maxTokens":15}, "operation": "completion" }' \
      http://localhost:<dapr-port>/v1.0/bindings/<binding-name>

Response

The response body contains the following JSON:

[
  {
    "finish_reason": "length",
    "index": 0,
    "text": " a pig in a dress.\n\nSun, Oct 20, 2013"
  },
  {
    "finish_reason": "length",
    "index": 1,
    "text": " the only thing on earth that loves you\n\nmore than he loves himself.\"\n\n"
  }
]

Chat Completion API

To perform a chat-completion operation, invoke the Azure OpenAI binding with a POST method and the following JSON body:

{
    "operation": "chat-completion",
    "data": {
        "deploymentId": "my-model",
        "messages": [
            {
                "role": "system",
                "message": "You are a bot that gives really short replies"
            },
            {
                "role": "user",
                "message": "Tell me a joke"
            }
        ],
        "n": 2,
        "maxTokens": 30,
        "temperature": 1.2
    }
}

The data parameters are:

• deploymentId - string that specifies the model deployment ID to use.
• messages - array of messages that will be used to generate chat completions. Each message is of the form:
  • role - string that specifies the role of the message. Can be either user, system or assistant.
  • message - string that specifies the conversation message for the role.
• maxTokens - (optional) defines the max number of tokens to generate. Defaults to 16 for the chat completion API.
• temperature - (optional) defines the sampling temperature between 0 and 2. Higher values like 0.8 make the output more random, while lower values like 0.2 make it more focused and deterministic. Defaults to 1.0 for the chat completion API.
• topP - (optional) defines the sampling temperature. Defaults to 1.0 for the chat completion API.
• n - (optional) defines the number of completions to generate. Defaults to 1 for the chat completion API.
• presencePenalty - (optional) Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far, increasing the model’s likelihood to talk about new topics. Defaults to 0.0 for the chat completion API.
• frequencyPenalty - (optional) Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model’s likelihood to repeat the same line verbatim. Defaults to 0.0 for the chat completion API.

Example

• Linux

curl -d '{
  "data": {
      "deploymentId": "my-model",
      "messages": [
          {
              "role": "system",
              "message": "You are a bot that gives really short replies"
          },
          {
              "role": "user",
              "message": "Tell me a joke"
          }
      ],
      "n": 2,
      "maxTokens": 30,
      "temperature": 1.2
  },
  "operation": "chat-completion"
}' \
http://localhost:<dapr-port>/v1.0/bindings/<binding-name>

Response

The response body contains the following JSON:

[
  {
    "finish_reason": "stop",
    "index": 0,
    "message": {
      "content": "Why was the math book sad? Because it had too many problems.",
      "role": "assistant"
    }
  },
  {
    "finish_reason": "stop",
    "index": 1,
    "message": {
      "content": "Why did the tomato turn red? Because it saw the salad dressing!",
      "role": "assistant"
    }
  }
]

Get Embedding API

The get-embedding operation returns a vector representation of a given input that can be easily consumed by machine learning models and other algorithms. To perform a get-embedding operation, invoke the Azure OpenAI binding with a POST method and the following JSON body:

{
    "operation": "get-embedding",
    "data": {
        "deploymentId": "my-model",
        "message": "The capital of France is Paris."
    }
}

The data parameters are:

• deploymentId - string that specifies the model deployment ID to use.
• message - string that specifies the text to embed.

Example

• Linux

curl -d '{
  "data": {
      "deploymentId": "embeddings",
      "message": "The capital of France is Paris."
  },
  "operation": "get-embedding"
}' \
http://localhost:<dapr-port>/v1.0/bindings/<binding-name>

Response

The response body contains the following JSON:

[0.018574921,-0.00023652936,-0.0057790717,.... (1536 floats total for ada)]

Related links

• Basic schema for a Dapr component
• Bindings building block
• How-To: Trigger application with input binding
• How-To: Use bindings to interface with external resources
• Bindings API reference
• Azure OpenAI Rest examples

Last modified October 12, 2023: Update config.toml (#3826) (0ffc2e7)