Question: [.Net] Is it possible to have ChatCompletionAgent connected to two or more data sources

[PanClifton]

hello
consider following C# code

 var chatAgent = new Microsoft.SemanticKernel.Agents.ChatCompletionAgent()
   {
       Description = "You are a friendly assisant for user",
       Instructions = configuration.Value.PromptConfiguration.SystemMessage,
       Kernel = _provideOpenAiKernel.Kernel,
       Arguments = new KernelArguments(new AzureOpenAIPromptExecutionSettings()
       {
           TopP = configuration.Value.ChatAgentConfiuration.TopP,
           Temperature = configuration.Value.ChatAgentConfiuration.Temperature,
           MaxTokens = configuration.Value.ChatAgentConfiuration.MaxTokens,
           AzureChatDataSource = dataSource, // <--- is it possible to have more than two datasources?
       }),
   };

is it possible to connect to agent more than one data source?

i have a business requirement that chat have to scan more than one data sources to generate a response and return citations.

i think i might have found solution for this problem.
Solution below is using one search client, obviously you could have more and merge data from them.

could you advise on following:
am i passing data from searchResult to agent via kernelArguments correctly?
how to retrieve citations, if possible using if (messageContext?.Citations is not null ...) in chat stream?

 [Fact]
 public async Task Test()
 {            
     var query = "Could you help me to prepare the project initiation phase by creating a checklist with activities and deliverables including their owners?";
     var searchClient1 = new SearchClient(
         new Uri(_aiSearchConfiguration.Endpoint),
         "general-knowledge",
         new AzureKeyCredential(_aiSearchConfiguration.ApiKey!));

     var options = new SearchOptions
     {
         Size = _aiSearchConfiguration.MaxSearchQueries,                                                               
     };

     var searchResultData = new List<string>();
     var citations = new List<string>();
             
     var searchResult = await searchClient1.SearchAsync<SearchDocument>(query, options);
     await foreach (var result in searchResult.Value.GetResultsAsync())
     {    
         var content = result.Document["content"]?.ToString() ?? "No content available.";
         var source = result.Document["source"]?.ToString() ?? "No content available.";
        
         searchResultData.Add(content);
         citations.Add(source);
     }

     var agent = new ChatCompletionAgent
     {
         Kernel = _fixture.ProvideOpenAiKernel.Kernel,
         Name = "MultiSourceAgent",
         Instructions = "Use the provided search results to generate a well-cited response."
     };

     var kernelArguments = new KernelArguments
     {
         { "context", searchResultData }, // am i correctly passing data to agent?
         { "citations", string.Join("\n", citations) }
     };

     var chatHistory = new ChatHistory();
     chatHistory.AddUserMessage(query);

     var responses = string.Empty;     

     await foreach (var content in agent.InvokeStreamingAsync(chatHistory, kernelArguments))
     {
         var contentReult = content.Content ?? string.Empty;

         responses = responses + contentReult;
         var update = content.InnerContent as OpenAI.Chat.StreamingChatCompletionUpdate;
         var messageContext = update?.GetMessageContext();

         if (messageContext?.Citations is not null)
         {
             // is it possible to get citations here?
         }
     }
 }

[crickman]

Hi @PanClifton, I suspect this question isn't specific to ChatCompletionAgent but is rather tied to the SK OpenAI chat-completion service and the underlying Azure services: https://github.com/Azure/azure-sdk-for-net/blob/main/sdk/openai/Azure.AI.OpenAI/tests/Samples/02_Oyd.cs

In short, I believe this is limited to one data-source.

You approach for using SearchClient directly seems a reasonable approach to transcend this limit.

KernelArguments are used to define values for parameters in the agent instructions.

'''
Generate a well cited response to the search results:

Results:
{{$context}}

Citations:
{{$citations}}
'''

Although, I might consider wrapping the search as a KernelFunction and then having the agent/model consume the results as part of a tool-call.