Returning output from a native function

[vicperdana]

Hi team,

Is there a way to get the output of native function's plugin. In this example below, I'd like to capture items when invoking kernel.

I'm using Auto Function Calling but can't find where the output results are stored. I mostly followed the example, with the addition of Auto Function Calling.

Would this be a case where a handlebar planner could fulfil the need?

[Description("Filter catalog items")]
    public class FilterCatalogItem (CatalogClientChatService catalogClientChatService)
    {
    // need to fix this code to ensure catalogServiceClient exists
        [KernelFunction, Description("Return a list of catalog items filtered by name or description")]
        public async Task<CatalogItemsPage> GetCatalogItems(
            [Description("Name or description of the item to be queried")] string? searchText)
        {
            if (catalogClientChatService is not null)
            {
                var items = await catalogClientChatService.SearchItemsAsync(searchText);
                return items;
            }
            else
            {
                throw new InvalidOperationException("CatalogServiceClient is not available");
            }
        }
    }

[dmytrostruk]

I'm using Auto Function Calling but can't find where the output results are stored.

@vicperdana Could you please provide an example where you execute auto function calling? If you are triggering auto function calling with IChatCompletionService, you should get a result for each function call in ChatHistory object.

[vicperdana]

Hi @dmytrostruk here's a snippet of the code. I am not using IChatCompletionService as I am streaming the results instead.

        OpenAIPromptExecutionSettings openAIPromptExecutionSettings = new()
        {
            MaxTokens = 200,
            ToolCallBehavior = ToolCallBehavior.AutoInvokeKernelFunctions
        };
       // More code...
       var result = _kernel.InvokeStreamingAsync<StreamingChatMessageContent>(
                  prompts["chat"],
                  new(openAIPromptExecutionSettings)
                  {
                      {"request", message},
                      {"history", string.Join("\n", history.Select(x => x.Role + ": " + x.Content)) }
                  }
                  );
              
        // Stream the response
        string resultMessage = "Aspire Shop: ";
        await foreach (var chunk in result)
        {
            if (chunk.Role.HasValue)
            {
                Console.Write(chunk.Role + " > ");
            }
            resultMessage += chunk;
            Console.Write(chunk);
        }
        
        history.AddUserMessage(message);
        history.AddAssistantMessage(resultMessage);

[dmytrostruk]

@vicperdana In this case, we do not provide function result, because streaming chunks are returning independently from function calling functionality. So, there are a couple of ways how to get each function result:

1. Instead of using Kernel, you can use IChatCompletionService (you can get it from kernel.GetRequiredService<IChatCompletionService>()) and execute streaming with GetStreamingChatMessageContentsAsync method. This method accepts ChatHistory object, and this object will be updated with each function result.
2. We are going to introduce filters for automatic function calling. Filters will allow you to intercept each function call, check function name, arguments, metadata and after function call you will be able to get function result. PR for this functionality and samples is here and it's going to be released soon: .Net: New API for Filters #5936.

Hope this information is helpful!