.Net: Bug: .NET Semantic Kernel Function Calling Not Working with Anthropic via Amazon Connector

[samitugal]

Describe the Bug

When I connect my Anthropic model via Microsoft.SemanticKernel.Connectors.Amazon to translate these sentences into English, the query processes I wrote as plugins — which normally work — are not receiving any requests.

However, when I perform the same operation with OpenAI services, the model can easily access these plugins. Unfortunately, this is not the case with Anthropic.

I'm experiencing the same issue with the latest o3-mini reasoning model. Could it be that the Amazon Connector is still in an alpha experimental phase, or that the OpenAI o3-mini version is too new, causing the function calling feature not to work properly?

Observations

When I send the same request to the gpt-4o-2024-08-06 model, the breakpoint it hits is:

However, when I send the same request to the anthropic.claude-3-5-sonnet-20240620-v1:0 model, it does not hit the breakpoint.

Implementation Details

The implementation for both clients is almost identical. The only difference is:

• For OpenAI services:

  • I use OpenAIPromptExecutionSettings and the AddOpenAIChatCompletion methods.
    

• For Anthropic services:

  • I use the base class PromptExecutionSettings because I couldn't find an implemented class specifically for AWS services.
    

---

Platform Information

• Language: C#

• NuGet Packages:

  • Microsoft.SemanticKernel: 1.37.0
  • Microsoft.SemanticKernel.Connectors.Amazon: 1.37.0-alpha
  • AWSBedrocRuntime: 4.0.0-preview.6

• AI Models:

  • Anthropic Model: anthropic.claude-3-5-sonnet-20240620-v1:0
  • OpenAI Model: gpt-4o-2024-08-06

• IDE: Microsoft Visual Studio Professional 2022 (3) (64-bit) - Current Version 17.13.0

• OS: Windows

---

Possible Causes

1. The Microsoft.SemanticKernel.Connectors.Amazon package is still in the alpha stage and might not fully support function calling.
2. The o3-mini model from OpenAI is very new, and compatibility issues might exist with the current function-calling implementation.
3. Missing AWS-specific PromptExecutionSettings class might be causing the model to bypass the plugin execution logic.

---

Additional Information

I'm happy to provide more details or logs if needed.
Thanks for your time and support!