Amazon - Bedrock

Amazon Bedrock is a service provided by Amazon Web Services (AWS) that allows you to access large language models with a serverless experience. Semantic Kernel provides a connector to access these models from AWS.

Prerequisites

An AWS account and access to the foundation models
AWS CLI installed and configured

Configuration

Follow this guide to configure your environment to use the Bedrock API.

Please configure the aws_access_key_id, aws_secret_access_key, and region otherwise you will need to create custom clients for the services. For example:

python

runtime_client=boto.client(
    "bedrock-runtime",
    aws_access_key_id="your_access_key",
    aws_secret_access_key="your_secret_key",
    region_name="your_region",
    [...other parameters you may need...]
)
client=boto.client(
    "bedrock",
    aws_access_key_id="your_access_key",
    aws_secret_access_key="your_secret_key",
    region_name="your_region",
    [...other parameters you may need...]
)

bedrock_chat_completion_service = BedrockChatCompletion(runtime_client=runtime_client, client=client)

Supports

Region

To find model supports by AWS regions, refer to this AWS documentation.

Inference profiles

You can create inference profiles in AWS Bedrock to monitor and optimize the performance of your foundation models. Refer to the AWS documentation for more information.

When you are using an Application Inference Profile, you must specify the BEDROCK_MODEL_PROVIDER environment variable to the model provider you are using. For example, if you are using Amazon Titan, you must set BEDROCK_MODEL_PROVIDER=amazon. This is because an Application Inference Profile doesn't contain the model provider information, and the Bedrock connector needs to know which model provider to use so that it can create the correct request body to the Bedrock API.

An Application Inference Profile ARN is usually formatted as followed: arn:aws:bedrock:<region>:<account-id>:application-inference-profile/<profile-id>.

Input & Output Modalities

Foundational models in Bedrock support the multiple modalities, including text, image, and embedding. However, not all models support the same modalities. Refer to the AWS documentation for more information.

The Bedrock connector supports all modalities except for image embeddings, and text to image.

Text completion vs chat completion

Some models in Bedrock supports only text completion, or only chat completion (aka Converse API), or both. Refer to the AWS documentation for more information.

Tool Use

Not all models in Bedrock support tools. Refer to the AWS documentation for more information.

Streaming vs Non-Streaming

Not all models in Bedrock support streaming. You can use the boto3 client to check if a model supports streaming. Refer to the AWS documentation and the Boto3 documentation for more information.

Model specific parameters

Foundation models can have specific parameters that are unique to the model or the model provider. You can refer to this AWS documentation for more information.

You can pass these parameters via the extension_data field in the PromptExecutionSettings object.

Unsupported features

Guardrail