Skip to content

AWS Bedrock Operations Guide

This guide outlines the current Bedrock validation surfaces that ship in this repository.

Prerequisites

  • Bedrock enabled in your AWS account and region (e.g., us-east-1).
  • Model access approved for the Claude family you intend to use (e.g., Sonnet/Haiku).
  • IAM user/role with permissions:
  • bedrock:InvokeModel
  • bedrock:InvokeModelWithResponseStream
  • Optional: sts:AssumeRole (CI usage)

Local Setup

# Choose a supported region
export AWS_REGION=us-east-1

# Use either a profile...
export AWS_PROFILE=traigent-bedrock
# ...or explicit keys (avoid in shared shells)
# export AWS_ACCESS_KEY_ID=...
# export AWS_SECRET_ACCESS_KEY=...

# Optional explicit model override
# export BEDROCK_MODEL_ID=anthropic.claude-3-sonnet-20240229-v1:0

# Local mock mode: export TRAIGENT_MOCK_LLM=true, or call
# traigent.testing.enable_mock_mode_for_quickstart() in tutorial code.
# Provider-specific *_MOCK env vars are ignored.

Credential Boundary

Bedrock credentials are AWS/customer credentials. The SDK's optional Bedrock client creates a local boto3/aioboto3 Bedrock Runtime client and relies on the AWS SDK credential chain on the machine where the optimization runs. The SDK does not use a Traigent API key as an AWS credential and does not send AWS credentials to the Traigent backend.

If you use AWS_BEARER_TOKEN_BEDROCK, treat it as a short-lived AWS/provider credential with the same boundary. It belongs in your local AWS credential environment or tooling, not in Traigent backend configuration.

For the broader credential and telemetry boundary, see the Credential & data trust model.

Dependency Installation

Ensure you installed integration dependencies: 

pip install -e ".[integrations]"  # or: pip install -r requirements/requirements-integrations.txt

Validation Checks

pytest -q tests/unit/integrations/test_cloud_plugins.py -o addopts='' -k bedrock
pytest -q tests/unit/integrations/test_bedrock_client.py -o addopts=''

Minimal Real Run

With AWS credentials configured, you can exercise the shipped Bedrock example:

python examples/integrations/bedrock/bedrock_hello.py

Bedrock coverage includes the native BedrockChatClient plus LangChain ChatBedrock and ChatBedrockConverse interception. In mock mode, TRAIGENT_MOCK_LLM routes Bedrock calls through MockAdapter; in real mode, responses are captured for token/cost/latency extraction.

Troubleshooting

  • AccessDeniedException: Verify model access is granted in Bedrock console and IAM policy includes InvokeModel*.
  • RegionError: Confirm AWS_REGION is a Bedrock-enabled region and matches your model access region.
  • ImportError: boto3: Run pip install boto3 botocore or install the integrations extra.

Notes

  • Costs: Start with small max tokens and low trial counts. Use AWS Budgets/Cost Anomaly detection alerts.
  • Reproducibility: Use RandomSamplerPlan to fix evaluation subsets across runs.