How to send KernelArguments towards KernelFunctions when using IChatCompletionService

[ystvan]

My goal is that the user can have a conversation with the model through Semantic Kernel's AzureOpenAIChatCompletion service.
I register two plugins as Kernel Functions that might be triggered during the conversation. I need to set the securityFilters value on runtime.

I cannot seem to find a way to initialize a Kernel Arguments arguments like

 var kernelArguments = new KernelArguments()
 {
     ["securityFilters"] = "some-string"
 };

An that could be used in a plugin as

[KernelFunction, Description("Provide document look up in company data for employees using SharePoint")]
[return: Description("document's filename, source as hyperlink and its content")]
public async Task<List<IndexSchema>> SearchAsync(
    [Description("The user intent in a a LUCENE search query format")] string query,
    [Description("The SharePoint security group memberships of the user")] string securityFilters)
{    
        return await _searchService.SearchAsync("index-01", query, securityFilter);
}

Where do I pass the KernelArguments into?

If it was a one-time prompt I could easily use InvokePromptStreamingAsync method that accepts KernelArguments, but I need the ChatCompletionService to maintain chat history, etc.

This is a rudimentary example where I struggle to pass in the KernelArguments :)

 var kernelBuilder = Kernel.CreateBuilder();

 kernelBuilder.AddAzureOpenAIChatCompletion(
          "model",
          "endpoint",
          "key");

 kernelBuilder.Plugins.AddFromType<AzureAISearchPlugin>();

 // Create kernel
 var kernel = kernelBuilder.Build();

 ChatHistory history = [];
 
 // Get chat completion service
 _chatCompletionService = kernel.GetRequiredService<IChatCompletionService>();

 // Enable auto function calling
 OpenAIPromptExecutionSettings openAIPromptExecutionSettings = new()
 {
     MaxTokens = 4000,
     ToolCallBehavior = ToolCallBehavior.AutoInvokeKernelFunctions
 };

 await foreach (var chatUpdate in _chatCompletionService.GetStreamingChatMessageContentsAsync(
                history,
                executionSettings: openAIPromptExecutionSettings,
                kernel: kernel))
 {
     yield return new TokenizedResponse(chatUpdate.Content);

     responseBuffer.Append(chatUpdate.Content);
 }

[dmytrostruk]

@ystvan Thank you for your question!

There are a couple of ways how to achieve your scenario:

1. If securityFilters property does not contain PII or any other sensitive information, in your chat history you can inject these details as part of user or system message. Since auto function invocation is enabled, LLM will be able to use that information and Semantic Kernel will inject it into your plugin.
2. If you do not wish to share securityFilters content with LLM, you could use kernel.Data property bag to set this information. In your plugin, you can inject kernel instance, which should contain that property, here is an example of Kernel injection:
   

   semantic-kernel/dotnet/samples/Concepts/Functions/MethodFunctions_Types.cs

   Lines 178 to 184 in 8d9c3cb

   	[KernelFunction]
   	public async Task<string> TaskInjectingKernelWithInputTextAndStringResult(Kernel kernel, string textToSummarize)
   	{
   	var summary = await kernel.InvokeAsync<string>(kernel.Plugins["SummarizePlugin"]["Summarize"], new() { ["input"] = textToSummarize });
   	this._output.WriteLine($"Running {nameof(this.TaskInjectingKernelWithInputTextAndStringResult)} -> Injected kernel + input: [textToSummarize: {textToSummarize[..15]}...{textToSummarize[^15..]}] -> result: {summary}");
   	return summary!;
   	}

Please let us know if any of these approaches work for you, thanks!

[ystvan]

Amazing! Did not know about the kernel.Data property, it came to the rescue! Thank you @dmytrostruk