This example shows streaming response from Amazon Bedrock with FastAPI on AWS Lambda.

This example uses Anthropic Claude v2 model to generate bedtime stories. FastAPI provides the static web frontend and an inference API. The inference API endpoint invokes Bedrock using Boto3, and streams the response. Both Lambda Web Adapter and function URL have response streaming mode enabled. So the response from Bedrock are streamed all the way back to the client.

This function is packaged as a Docker image. Here is the content of the Dockerfile.

FROM public.ecr.aws/docker/library/python:3.12.0-slim-bullseye
COPY --from=public.ecr.aws/awsguru/aws-lambda-adapter:0.8.4 /lambda-adapter /opt/extensions/lambda-adapter

WORKDIR /app
ADD . .
RUN pip install -r requirements.txt

CMD ["python", "main.py"]

Notice that we only need to add the second line to install Lambda Web Adapter.

COPY --from=public.ecr.aws/awsguru/aws-lambda-adapter:0.8.4 /lambda-adapter /opt/extensions/

In the SAM template, we use an environment variable AWS_LWA_INVOKE_MODE: RESPONSE_STREAM to configure Lambda Web Adapter in response streaming mode. And adding a function url with InvokeMode: RESPONSE_STREAM.

  FastAPIFunction:
    Type: AWS::Serverless::Function
    Properties:
      PackageType: Image
      MemorySize: 512
      Environment:
        Variables:
          AWS_LWA_INVOKE_MODE: RESPONSE_STREAM
      FunctionUrlConfig:
        AuthType: NONE
        InvokeMode: RESPONSE_STREAM
      Policies:
      - Statement:
        - Sid: BedrockInvokePolicy
          Effect: Allow
          Action:
          - bedrock:InvokeModelWithResponseStream
          Resource: '*'

Run the following commends to build and deploy this example.

sam build --use-container
sam deploy --guided

After the deployment completes, open the FastAPIFunctionUrl shown in the output messages. You should see a simple web page. Here is a demo.