Amazon Bedrock

Region-bound Bedrock Runtime client with Anthropic InvokeModel and native Converse or ConverseStream routes.

Contract and endpoint

The public constructor is func New(creds auth.SigV4Credentials, region string, options ...Option) (inference.Client, error). It binds provider identity bedrock and resolves the base below when Model.BaseURL is empty, unless the dynamic rule says otherwise.

New requires region, access key, and secret key. Optional session tokens are signed. OpenAI format is rejected before I/O.

An explicit model BaseURL is caller-controlled and is used by the provider route builder. It replaces the package default; it is not appended to the default.

Authentication and model formats

The llm provider registry classifies this identity as requiring AWS SigV4. The constructor receives resolved credential material rather than a secret reference. A credentials or secrets integration can resolve a descriptor and lease before passing the value here; secret labels do not belong in model.Model.

Optional capability bits on model.Model remain caller input. Request validation is authoritative for tools, structured output, images, thinking, sampling, and output limits; this page records only features encoded by the selected codec or provider patch.

Streaming, structured output, and tools

The client returns the shared stream reader from Stream. Its selected codec encodes provider-neutral tool declarations, tool results, and structured-output schemas when the request is valid. Streaming responses preserve codec usage and finish metadata; no capability beyond the source-backed format is inferred.

A stream owns its response body through the returned reader. Transport and non-success HTTP failures are returned before a reader is exposed; decode failures surface from Next or the terminal result.

%%{init: {"theme":"base","themeVariables":{"background":"#0b1020","primaryColor":"#172554","primaryTextColor":"#f8fafc","primaryBorderColor":"#60a5fa","lineColor":"#94a3b8","secondaryColor":"#1e293b","tertiaryColor":"#111827","fontFamily":"ui-sans-serif,system-ui"}}}%%
flowchart LR
    Req[Request] --> Enc[format codec]
    Enc --> Wire[provider route]
    Wire --> Stream[stream framing]
    Stream --> Reader[StreamReader]
    Reader --> Result[terminal result and usage]
    classDef dark fill:#172554,stroke:#60a5fa,color:#f8fafc;
    class Req,Enc,Wire,Stream,Reader,Result dark;

Caching controls

WithPromptCachePoint adds a native Converse cache point with the declared TTL and does not alter the Anthropic InvokeModel path.

Counters, errors, and retries

NewCounter sends exact countTokens requests for supported Anthropic and Converse formats; unsupported formats return UnsupportedAPIFormatError.

The shared inference boundary returns typed model-validation, authentication, network, HTTP, request-encoding, response-decoding, and stream errors. A provider-local retry loop is not implied by a constructor name.

No provider-local response retry is declared. Use inference/retry only around an operation your caller can repeat safely.

Consumer example

The example keeps credentials out of model.Model and calls the public constructor. Replace environment lookup with an application-owned credentials or secrets lease.

package example
import (
    "context"
    "os"
    "github.com/looprig/inference"
    "github.com/looprig/inference/auth"
    model "github.com/looprig/inference/model"
    bedrock "github.com/looprig/llm/providers/bedrock"
)
func invoke() error {
    // Resolve credentials from an application-owned lease or environment, never model.Model.
    selected := model.CustomModel(model.ProviderName("bedrock"), model.APIFormatBedrockConverse, "", "model-name")
    _ = selected
    // Construction validates provider policy and binds its endpoint and codec.
    client, err := bedrock.New(auth.SigV4Credentials{AccessKeyID: os.Getenv("AWS_ACCESS_KEY_ID"), SecretAccessKey: os.Getenv("AWS_SECRET_ACCESS_KEY")}, "us-east-1")
    if err != nil { return err }
    // The context controls the request lifetime, including a stream if used.
    _, err = client.Invoke(context.Background(), inference.Request{Model: selected})
    return err
}

Source and proof

• body.go
• client.go
• counter.go
• errors.go
• options.go

The provider identity and API-format truth table are defined in provider.go. Adjacent behavior tests:

• body_test.go
• client_test.go
• converse_test.go
• counter_test.go
• export_test.go
• options_test.go
• usage_test.go