I am experimenting with differnt agent patterns in dotnet.
Lets compare different options for function calling with Amazon Bedrock:
- AWSSDK.BedrockRuntime
- AgentDo
- Microsoft.Extensions.AI & AWSSDK.Extensions.Bedrock.MEAI
- Microsoft.SemanticKernel & Microsoft.SemanticKernel.Connectors.Amazon
Warning
With the advent of Microsoft Agent Framework (building on MEAI), Semantic Kernel does not have a future!
There seems to be an issue regarding #11448 .Net: Support Function calling for Amazon Bedrock (Using Converse APIs). The recommendation seems to be to use IChatClient instead of IChatCompletionService. And I actually got function calling working with like this:
var kernel = serviceScope.ServiceProvider.GetRequiredService<Kernel>();
var chatClient = kernel.GetRequiredService<IChatClient>();
List<ChatMessage> chatHistory = [];
chatHistory.Add(new ChatMessage(ChatRole.System, "You are a helpful AI assistant"));
chatHistory.Add(new ChatMessage(ChatRole.User, "Do I need an umbrella?"));
var invocation = chatClient.GetStreamingResponseAsync(
messages: chatHistory,
options: new()
{
Temperature = 0f,
Tools = [AIFunctionFactory.Create([Description("Get the current weather.")]() => kernel.GetRequiredService<WeatherInformation>().GetWeather())]
});
await foreach (var update in invocation)
{
Console.Write(update);
}However it feels less well integrated:
ChatHistoryis not updated automatically, but messages need to be maintained manually.- Tools are not picked up automatically from plugins but have to be provided again.
There seems to be interoperability method available with chatClient.AsChatCompletionService();, but this breaks function calling again.
It seems that with IChatCompletionService my tools / plugins do not get picked up when used with AWS Bedrock.
Bedrock is also not on the offical list of function calling support:
According to AWS Bedrock anthropic claude tool call integration with microsoft semantic kernel a custom implementation can adapt the AWSSDK.BedrockRuntime IChatClient, like resources/2025-04-02/ConsoleApp/ConsoleApp/AnthropicChatClient.cs. The implementation reminds me of AgentDo, but seems more comprehensive.
It is set up like this:
services.AddKeyedSingleton<IChatCompletionService>("anthropic.adapter", (sp, key) =>
{
var runtime = sp.GetRequiredService<IAmazonBedrockRuntime>();
IChatClient client = new AnthropicChatClient(runtime, "anthropic.claude-3-5-sonnet-20240620-v1:0");
#pragma warning disable SKEXP0001 // Type is for evaluation purposes only and is subject to change or removal in future updates. Suppress this diagnostic to proceed.
IChatCompletionService chatCompletionService = client
.AsBuilder()
.UseFunctionInvocation()
.Build()
.AsChatCompletionService();
#pragma warning restore SKEXP0001 // Type is for evaluation purposes only and is subject to change or removal in future updates. Suppress this diagnostic to proceed.
return chatCompletionService;
});It successfully invokes my tools / plugins.