The integration of chat and text completion services into the context does not utilize the models as expected #3103

Guillermo Subirán (gsubiran) · 2023-10-09T14:24:43Z

Guillermo Subirán (gsubiran)
Oct 9, 2023

I'm developing an Azure Function and I've integrated SK to it injecting the kernel as seen in many starter projects.

//Microsoft Semantic Kernel configuration build
var skBuild = Kernel.Builder
                    .WithLoggerFactory(loggerFactory)
                    .WithAzureTextEmbeddingGenerationService("text-embedding-ada-002", azureOpenAIOptions.Endpoint, azureOpenAIOptions.ApiKey)
                    .WithAzureTextCompletionService("gpt-35-turbo-instruct", azureOpenAIOptions.Endpoint, azureOpenAIOptions.ApiKey)
                    .WithAzureChatCompletionService("gpt-35-turbo", azureOpenAIOptions.Endpoint, azureOpenAIOptions.ApiKey)                        
                    .WithMemoryStorage(memoryStore)
                    .Build();

Even I specified different models for each searvice

.WithAzureTextCompletionService("gpt-35-turbo-instruct",... ,...)

and

.WithAzureChatCompletionService("gpt-35-turbo",... ,...)

SK is only using the model specified for chat completion service.
What I'm expecting is to SK uses the correct model depending on implementation.

For example:
gpt-35-turbo-instruct when implementing

var result = await kernel.RunAsync(context, skill["Joke"]);
    Console.WriteLine(result);

and gpt-35-turbo when implementing

var chat = chatCompletionService.CreateNewChat("You are an AI assistant that helps people find information.");
        chat.AddMessage(AuthorRole.User, "Hi, what information can you provide for me?");
    
        string response = await chatCompletionService.GenerateMessageAsync(chat, new ChatRequestSettings());
        Console.WriteLine(response);