You are viewing the docs for Archestra v1.4. View the latest docs
Amazon Bedrock
Use every model in your Amazon Bedrock account through Archestra. Connect with a Bedrock API key or AWS access keys. On AWS, Archestra can use its own IAM role, so it stores no key.
Adding a Provider
- Go to Model Providers → Add API Key and select Amazon Bedrock.
- Choose the authentication method:
- API Key: paste a Bedrock API key.
- AWS SigV4: enter Access Key ID, Secret Access Key, and an optional Session Token.
- Service Account: use the IAM credentials available to the backend. This requires
ARCHESTRA_BEDROCK_IAM_AUTH_ENABLED=true.
- Choose the Region where your models are available.
- Click Test & Create.
The key appears in the provider table, and available models appear under Models. For a private endpoint, set Base URL under Advanced. Ensure the selected region matches that endpoint.
Proxy clients use https://<archestra-host>/v1/bedrock. The proxy supports Bedrock Converse and InvokeModel requests. Through Model Router, use bedrock:<model-id>.
IAM Authentication Setup (IRSA)
To use IAM authentication on EKS with IRSA:
-
Create an IAM role with
AmazonBedrockFullAccessor a scoped policy (see below) -
Create an OIDC provider for your EKS cluster
-
Configure the IAM role's trust policy to allow the Archestra service account:
json{ "Effect": "Allow", "Principal": { "Federated": "arn:aws:iam::<ACCOUNT_ID>:oidc-provider/oidc.eks.<REGION>.amazonaws.com/id/<OIDC_ID>" }, "Action": "sts:AssumeRoleWithWebIdentity", "Condition": { "StringEquals": { "oidc.eks.<REGION>.amazonaws.com/id/<OIDC_ID>:sub": "system:serviceaccount:archestra:archestra-platform" } } } -
Annotate the Archestra service account:
bashkubectl annotate sa archestra-platform -n archestra \ eks.amazonaws.com/role-arn=arn:aws:iam::<ACCOUNT_ID>:role/<ROLE_NAME> -
Set
ARCHESTRA_BEDROCK_IAM_AUTH_ENABLED=trueand restart the deployment. Create the provider key using Service Account.
Minimum IAM Policy
Archestra calls the Bedrock Converse API, and the InvokeModel API for clients that use it (Claude Code, for example). Converse uses the same inference permissions as InvokeModel; see the AWS Converse API reference. The IAM role needs these actions:
{
"Version": "2012-10-17",
"Statement": [
{
"Effect": "Allow",
"Action": [
"bedrock:InvokeModel",
"bedrock:InvokeModelWithResponseStream"
],
"Resource": [
"arn:aws:bedrock:*:<ACCOUNT_ID>:inference-profile/us.anthropic.*",
"arn:aws:bedrock:*::foundation-model/anthropic.*"
]
},
{
"Effect": "Allow",
"Action": [
"bedrock:ListInferenceProfiles",
"bedrock:ListFoundationModels"
],
"Resource": "*"
}
]
}
Use * for the region in resource ARNs — cross-region inference profiles (us. prefix) can route requests to any US region.
The two list actions populate the model picker. ListInferenceProfiles returns cross-region and application inference profiles. ListFoundationModels adds on-demand models that have no inference profile. Without it, those models are not offered.
Model Discovery
Set ARCHESTRA_ to a comma-separated list such as anthropic,amazon to limit discovered model vendors. ARCHESTRA_ limits inference-region prefixes such as us,global. Empty values allow all.
Structured Output
Get JSON that matches your schema from a Bedrock model through the OpenAI-compatible endpoints. Send response_ with type: "json_schema" to /v1/bedrock/openai/chat/completions or the Model Router. Archestra passes the schema to Bedrock as a native structured output constraint.
What to know:
- Bedrock enforces a schema only on models that support structured output. For other models, Archestra sends the request without the schema, and the answer is not constrained.
- Bedrock has no JSON mode. Archestra accepts
type: "json_object"and sends the request without a constraint, so ask for JSON in the prompt. - Bedrock returns a 400 error for a schema it cannot use. The AWS page lists the supported JSON Schema features.
Prompt Caching
The native Converse and ConverseStream endpoints forward caller-supplied cachePoint blocks in system messages, message content, and tool definitions. Cache reads and writes reported by Bedrock are recorded.
Archestra does not insert checkpoints into native proxy requests. Use a model and region that support Bedrock prompt caching; checkpoints do not guarantee cache hits.
Through the Model Router, use cache_control: { "type": "ephemeral" } on supported OpenAI-shaped content blocks, messages, or function tool definitions. The router converts these explicit boundaries to Bedrock cachePoint blocks for Chat Completions and Responses requests; it does not create checkpoints automatically.
Top-level cache_ is not a Bedrock checkpoint. The router rejects it, incompatible provider cache options, and conflicting outer and inner function-tool cache markers with a 400 error.