Universal.Operative.Sdk.AwsBedrock 1.0.0

About

Universal.Operative.Sdk.AwsBedrock plugs Amazon Bedrock's Converse API into Universal.Operative.Sdk via an IModel adapter (Model) backed by the official AWSSDK.BedrockRuntime client.

Converse/ConverseStream present one provider-neutral request/response shape across every Bedrock model family that implements it — Anthropic Claude, Meta Llama, Mistral, Amazon Nova/Titan, Cohere Command, AI21 Jamba, DeepSeek, and more — so this single adapter works with any Bedrock model or cross-region inference profile. Point ModelId at whichever one you want; there's no per-family adapter to pick.

How to Use

Installation

dotnet add package Universal.Operative.Sdk.AwsBedrock

Configuration

using Universal.Operative.Sdk;
using Universal.Operative.Sdk.AwsBedrock;

// Magic credentials: resolves from the environment (AWS_ACCESS_KEY_ID /
// AWS_SECRET_ACCESS_KEY / AWS_SESSION_TOKEN or AWS_REGION), the shared
// credentials/config file and profile, container/instance role, SSO, etc. —
// the same default chain the AWS CLI and every other AWS SDK use.
IModel model = new Model(new Model.Options
{
    ModelId = "anthropic.claude-sonnet-4-5-20250929-v1:0",
    MaxTokens = 4096, // required by Bedrock's Converse API; defaults to 1024
});

Pass explicit credentials and/or a region when you don't want the default chain:

using Amazon.Runtime;

IModel model = new Model(new Model.Options
{
    ModelId = "us.anthropic.claude-sonnet-4-5-20250929-v1:0", // cross-region inference profile
    Region = "us-east-1",
    Credentials = new BasicAWSCredentials(accessKeyId, secretAccessKey),
});

The constructor also accepts an existing Amazon.BedrockRuntime.IAmazonBedrockRuntime in place of Options.Region/Credentials, so you can reuse a pre-built client (custom retry policy, endpoint override, assumed-role credentials, etc.) across models:

using Amazon;
using Amazon.BedrockRuntime;

IAmazonBedrockRuntime client = new AmazonBedrockRuntimeClient(RegionEndpoint.USEast1);
IModel model = new Model(client, new Model.Options { ModelId = "meta.llama3-1-70b-instruct-v1:0" });

Usage

Drop the model into OperativeBuilder like any other IModel:

IOperative operative = OperativeBuilder.FromDefaults(model)
    .AddBaselineTools()
    .Build();

await foreach (OperativeEvent e in operative.ExecuteAsync(conversation))
{
    // handle events as usual — see Universal.Operative.Sdk's Package.md
}

Any Bedrock model

ModelId is passed through verbatim — set it to a plain model id ("anthropic.claude-3-5-sonnet-20241022-v2:0", "meta.llama3-1-70b-instruct-v1:0", "amazon.nova-pro-v1:0", "mistral.mistral-large-2407-v1:0", "cohere.command-r-plus-v1:0") or a region-prefixed cross-region inference profile id ("us.anthropic.claude-sonnet-4-5-20250929-v1:0"). Features Converse doesn't support for a given model (tool use, images, extended reasoning, ...) surface as a Bedrock ValidationException at call time — this adapter doesn't pre-validate per-model capability.

Model-specific extras

Model.Options.AdditionalModelRequestFields forwards raw JSON verbatim as Bedrock's additionalModelRequestFields — the escape hatch for knobs specific to one model family (e.g. Anthropic's thinking block when calling Claude on Bedrock) that the universal IModel surface doesn't model.

Prompt caching

Set UsePromptCaching = true to stamp a cache checkpoint on the system prompt, the tool definitions, and the last message, so repeated turns and a stable system prompt/tool set are served from Bedrock's prompt cache:

IModel model = new Model(new Model.Options
{
    ModelId = "anthropic.claude-sonnet-4-5-20250929-v1:0",
    MaxTokens = 4096,
    UsePromptCaching = true,
});

Unlike the Anthropic adapter, this defaults to false — Bedrock caching support varies by model family, and enabling it for a model that doesn't support it returns a ValidationException. Cache read/write token counts round-trip on Usage.Bedrock.CacheReadInputTokens / CacheWriteInputTokens; pass AwsBedrock.BedrockUsageExtensions.ProjectCompactionTokens as CompactionTransformer.Options.ProjectedTokensCalculator so auto-compaction accounts for them (they're additive on top of Usage.InputTokens, not included in it — the same accounting shape as Anthropic's cache counters).

License

Free for noncommercial use under the Universal.Operative Noncommercial License. Source is closed; commercial use requires a separate license from Andrew Ong.