stache-ai-bedrock

Amazon Bedrock provider for Stache AI - LLM and embedding support via AWS Bedrock.

Installation

pip install stache-ai-bedrock

Usage

Install the package and configure the provider in your settings:

from stache_ai.config import Settings

settings = Settings(
    llm_provider="bedrock",
    embedding_provider="bedrock",
    bedrock_model_id="anthropic.claude-sonnet-4-20250514-v1:0",  # Optional
    bedrock_embedding_model_id="cohere.embed-english-v3",        # Optional
)

The provider will be automatically discovered via entry points.

Environment Variables

Supported Models

LLM Models

The provider supports all Bedrock-available models including:

Embedding Models

IAM Permissions

The Bedrock provider requires specific IAM permissions for model invocation.

Minimum Required Permissions

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream",
        "bedrock:Converse"
      ],
      "Resource": "*"
    },
    {
      "Effect": "Allow",
      "Action": [
        "aws-marketplace:ViewSubscriptions",
        "aws-marketplace:Subscribe"
      ],
      "Resource": "*"
    }
  ]
}

Why Resource: "*"?

Bedrock uses cross-region inference profiles for some models, which require wildcard resource permissions. Specific model ARNs (e.g., arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-*) will fail for cross-region requests.

SAM Template Example

Resources:
  StacheFunction:
    Type: AWS::Serverless::Function
    Properties:
      Environment:
        Variables:
          LLM_PROVIDER: bedrock
          EMBEDDING_PROVIDER: bedrock
          BEDROCK_MODEL_ID: anthropic.claude-sonnet-4-20250514-v1:0
          BEDROCK_EMBEDDING_MODEL_ID: cohere.embed-english-v3
      Policies:
        - Version: '2012-10-17'
          Statement:
            - Effect: Allow
              Action:
                - bedrock:InvokeModel
                - bedrock:InvokeModelWithResponseStream
                - bedrock:Converse
              Resource: '*'
            - Effect: Allow
              Action:
                - aws-marketplace:ViewSubscriptions
                - aws-marketplace:Subscribe
              Resource: '*'

Terraform Example

data "aws_iam_policy_document" "bedrock" {
  statement {
    effect = "Allow"
    actions = [
      "bedrock:InvokeModel",
      "bedrock:InvokeModelWithResponseStream",
      "bedrock:Converse"
    ]
    resources = ["*"]
  }

  statement {
    effect = "Allow"
    actions = [
      "aws-marketplace:ViewSubscriptions",
      "aws-marketplace:Subscribe"
    ]
    resources = ["*"]
  }
}

resource "aws_iam_role_policy" "bedrock" {
  name   = "bedrock-access"
  role   = aws_iam_role.lambda.id
  policy = data.aws_iam_policy_document.bedrock.json
}

Important Notes

Model Access

Before using a model, you must enable access in the AWS Bedrock console:

1. Go to AWS Console → Bedrock → Model access
2. Request access to the models you want to use
3. Wait for access approval (usually instant for most models)

Converse API

This provider uses the Bedrock Converse API (not the legacy InvokeModel API for chat). This requires the bedrock:Converse permission - bedrock:InvokeModel alone is not sufficient for LLM chat operations.

Embeddings

Embeddings use bedrock:InvokeModel with the embedding model. The Cohere embed models return 1024-dimensional vectors by default.

Requirements

• Python >= 3.10
• stache-ai >= 0.1.0
• boto3 >= 1.34.0