Text Generation and Chat Completion

[charlizeyeh]

Add Amazon Bedrock Connector

This pull request creates a connector for Amazon Bedrock, expanding the Semantic Kernel's integration capabilities with Bedrock's many foundation models.

Key Features:

• Implements IChatCompletionService and ITextGenerationService from the Semantic Kernel
• Includes client authentication and secure API communication with Amazon Bedrock
• Has core, client, service, and extension folders as directory organization and namespace similar to other connectors
• Includes unit tests and integration tests for available models

Implementation Details:

• Amazon Bedrock Runtime actions InvokeModelAsync and InvokeModelWithResponseStreamAsync were used for GetTextContentsAsync and GetStreamingTextContentsAsync respectively

• Amazon Bedrock Runtime actions ConverseAsync and ConverseStreamAsync were used for GetChatMessageContentsAsync and GetStreamingChatMessageContentsAsync respectively

• All versions of the following base models are currently connected and available for use under Bedrock:

  • AI21 Labs Jamba
  • AI 21 Labs Jurassic-2 series
  • Amazon Titan
  • Anthropic Claude
  • Cohere Command
  • Meta Llama 3
  • Mistral AI Mistral

• The following authentication scenarios are supported by the kernel builder extension methods:

  • User provides the IAmazonBedrockRuntime object and model ID: This approach gives the user more control over the configuration of the IAmazonBedrockRuntime object, such as specifying custom credentials, endpoints, or other settings.
  • User only provides model ID: An instance of IAmazonBedrockRuntime is automatically created and registered in the dependency injection (DI) container using the AWSSDK.Extensions.NETCore.Setup package, which adds AWS service clients to the DI container. In this case, the IAmazonBedrockRuntime object will be configured with default settings, such as credentials obtained from the environment variables, AWS credentials file, or instance profile credentials.

Testing:

• Unit tests cover core functionality and error handling
• Integration tests demonstrate successful communication with Bedrock API, request/response mapping, and output generation

[Beau-Gosse-dev]

I still need to review the implementation specific to each model/IOService

[Beau-Gosse-dev]

Okay, these are my last comments. Looks really good overall though, great job!

[Beau-Gosse-dev]

Main review is now here: microsoft#7587