Azure OpenAI details are not recognized

[marton-eifert]

Hello, I was trying out the 0.3.0 version on my local (Windows) laptop using R 4.3.1 in RStudio.
Since I have an Azure OpenAI Services deployment running gpt-35-turbo, I wanted to set the env variables according to descriptions in the README.
I changed my (project scope and user scope) .Renviron file accordingly and it contains the following lines now (entries masked):

AZURE_OPENAI_TASK="azure"
AZURE_OPENAI_ENDPOINT="<my_endpoint_url>"
AZURE_OPENAI_DEPLOYMENT_NAME="<my_deloyment_name>"
AZURE_OPENAI_KEY="<my_api_endpoint_key>"
AZURE_OPENAI_API_VERSION="2023-07-01-preview"

However, it does not recognize the variables, it still asks me for the OPENAI_API_KEY environment variable which I did not provide because I don't have one and I do want to use my Azure endpoint.

One other thing that is missing in the README --> What is the "TASK" supposed to be? I only found something called AZURE_OPENAI_TYPE in the context of the Microsoft Quickstart guide and "openai Python library" -- which I, by the way, have successfully set up with my AZURE_OPENAI_xxx environment variables elsewhere already --.
That one is referring to the API_TYPE which in the case of Azure OpenAI Services should take the value "azure" as above.
So is TYPE = TASK? Or is it something else?

[chendakeng]

Yes, I have the same problems as you. I am googling all over and cannot find anything related to "task". I hope the team can give us a clear instruction with examples of what to fill in for these environment variables.

[calderonsamuel]

Inspecting the source code shows that the base request for the azure openai method is defined as the following.

request_base_azure_openai <-
  function(task = Sys.getenv("AZURE_OPENAI_TASK"),
           base_url = Sys.getenv("AZURE_OPENAI_ENDPOINT"),
           deployment_name = Sys.getenv("AZURE_OPENAI_DEPLOYMENT_NAME"),
           token = Sys.getenv("AZURE_OPENAI_KEY"),
           api_version = Sys.getenv("AZURE_OPENAI_API_VERSION")
  ) {
    httr2::request(base_url) %>%
      httr2::req_url_path_append("openai/deployments") %>%
      httr2::req_url_path_append(deployment_name) %>%
      httr2::req_url_path_append(task) %>%
      httr2::req_url_path_append(api_version) %>%
      httr2::req_headers("api-key" = token,
                         "Content-Type" = "application/json") %>%
      httr2::req_method("POST")
  }

So this is how the request is translated. The task gets appended to path after the model name.

request_base_azure_openai(
  base_url = "https://www.example.com", 
  deployment_name = "gpt-35-turbo-instruct", 
  task = "completions", 
  api_version = "2023-05-15",
  token = "MY_API_KEY"
)
#> <httr2_request>
#> POST
#> https://www.example.com/openai/deployments/gpt-35-turbo-instruct/completions/2023-05-15
#> Headers:
#> • api-key: 'MY_API_KEY'
#> • Content-Type: 'application/json'
#> Body: empty

By looking at the REST documentation, it looks to me like you can just set it as "completions".

AZURE_OPENAI_TASK="completions"

The same documentation tells us that there is still a small bug that will prevent it from working. This should be how the base request is defined, with the api version as a url query parameter:

request_base_azure_openai <-
  function(task = Sys.getenv("AZURE_OPENAI_TASK"),
           base_url = Sys.getenv("AZURE_OPENAI_ENDPOINT"),
           deployment_name = Sys.getenv("AZURE_OPENAI_DEPLOYMENT_NAME"),
           token = Sys.getenv("AZURE_OPENAI_KEY"),
           api_version = Sys.getenv("AZURE_OPENAI_API_VERSION")
  ) {
    httr2::request(base_url) %>%
      httr2::req_url_path_append("openai/deployments") %>%
      httr2::req_url_path_append(deployment_name) %>%
      httr2::req_url_path_append(task) %>%
      httr2::req_url_query("api-version" = api_version) %>%
      httr2::req_headers("api-key" = token,
                         "Content-Type" = "application/json") %>%
      httr2::req_method("POST")
  }

request_base_azure_openai(
  base_url = "https://www.example.com", 
  deployment_name = "gpt-35-turbo-instruct", 
  task = "completions", 
  api_version = "2023-05-15",
  token = "MY_API_KEY"
)
#> <httr2_request>
#> POST
#> https://www.example.com/openai/deployments/gpt-35-turbo-instruct/completions?api-version=2023-05-15
#> Headers:
#> • api-key: 'MY_API_KEY'
#> • Content-Type: 'application/json'
#> Body: empty

I'll make a PR about this

^{Created on 2023-11-02 with reprex v2.0.2}