Compendium.Adapters.Bedrock 1.0.0-preview.0

Compendium.Adapters.Bedrock

AWS Bedrock implementation of Compendium's IAIProvider — one adapter, every hosted model family : Anthropic Claude, Meta Llama, Mistral, Amazon Nova, Amazon Titan, Cohere. Text completions (sync + streaming) flow through Bedrock's unified Converse / ConverseStream API; embeddings flow through InvokeModel with per-family payload shapes.

This adapter is for callers who want AWS data-residency, IAM-based access control, SigV4 auth, and Amazon's commercial terms — not the direct Anthropic API (see Compendium.Adapters.Anthropic for that).

Quick start

using Compendium.Abstractions.AI;
using Compendium.Abstractions.AI.Models;
using Compendium.Adapters.Bedrock.DependencyInjection;
using Microsoft.Extensions.DependencyInjection;

var services = new ServiceCollection();
services.AddLogging();
services.AddCompendiumBedrock(o =>
{
    o.Region = "us-east-1";
    o.DefaultModelId = "anthropic.claude-3-5-sonnet-20241022-v2:0";
    o.EmbeddingModelId = "amazon.titan-embed-text-v2:0";
    // Credentials default to the AWS SDK chain (env, profile, IAM role).
});

var sp = services.BuildServiceProvider();
var ai = sp.GetRequiredService<IAIProvider>();

var result = await ai.CompleteAsync(new CompletionRequest
{
    Model = "anthropic.claude-3-haiku-20240307-v1:0",
    Messages = new List<Message> { Message.User("Hello, Bedrock.") },
});

Console.WriteLine(result.Value.Content);

Run the sample :

AWS_REGION=us-east-1 dotnet run --project samples/01-chat-with-claude-via-bedrock

IAM policy

The runtime adapter needs only data-plane permissions :

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "BedrockRuntime",
      "Effect": "Allow",
      "Action": [
        "bedrock:Converse",
        "bedrock:ConverseStream",
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream"
      ],
      "Resource": [
        "arn:aws:bedrock:*::foundation-model/*",
        "arn:aws:bedrock:*:*:inference-profile/*"
      ]
    }
  ]
}

Scope Resource down to specific model ARNs in production. The adapter does not require bedrock:ListFoundationModels — the model catalog is curated in-package.

Options

Bind from IConfiguration :

{
  "Compendium": {
    "Adapters": {
      "Bedrock": {
        "Region": "us-east-1",
        "DefaultModelId": "anthropic.claude-3-5-sonnet-20241022-v2:0",
        "EmbeddingModelId": "amazon.titan-embed-text-v2:0"
      }
    }
  }
}

services.AddCompendiumBedrock(Configuration);

Model id routing

Bedrock's Converse API is family-agnostic — the adapter passes the model id through as-is. Use the canonical Bedrock id format :

• anthropic.claude-3-5-sonnet-20241022-v2:0
• anthropic.claude-3-haiku-20240307-v1:0 (cheapest Claude — great default for dev/CI)
• meta.llama3-1-70b-instruct-v1:0
• mistral.mistral-large-2407-v1:0
• amazon.nova-pro-v1:0

Cross-region inference profiles work too :

• us.anthropic.claude-3-5-sonnet-20241022-v2:0
• eu.anthropic.claude-3-5-sonnet-20241022-v2:0

Embeddings auto-detect the family by id prefix :

• amazon.titan-embed-* → Titan single-text per call, supports the dimensions hint.
• cohere.embed-* → Cohere batch (one HTTP call for the whole input list), defaults to input_type = "search_document".

Unknown embedding ids fail fast with AI.InvalidRequest.

Billing & throughput

Bedrock offers two billing modes per model :

1. On-demand — pay per token. The default; just call the adapter.
2. Provisioned throughput — pay for reserved capacity by the hour. Set DefaultModelId to the provisioned model ARN (arn:aws:bedrock:...:provisioned-model/...) and the adapter routes there.

Prompt caching (currently Anthropic-only on Bedrock) is not yet exposed by this preview; it will land behind a BedrockOptions.EnablePromptCaching flag in a follow-up.

Data residency

All requests stay within the configured Region. Bedrock does not train on your prompts. For workloads with strict residency requirements, pick a region in the appropriate AWS partition (eu-central-1, ap-northeast-1, GovCloud, ...) and confirm model availability with aws bedrock list-foundation-models.

Production checklist

• IAM role, not static access keys — use EKS pod identity or EC2 instance profile.
• Region pinning — set Region explicitly; do not rely on the AWS SDK's AWS_REGION env var alone in multi-region deployments.
• Scope IAM resources to specific model ARNs.
• Set MaxRetries in line with your latency SLO — Bedrock returns 429 under load.
• Log via ILogger<BedrockAIProvider> — turn it on for failed-request diagnostics.
• Hook a health check — HealthCheckAsync() sends a 1-token probe.
• Cap DefaultMaxTokens to avoid runaway costs from misbehaving callers.
• Pin the model id — Bedrock occasionally retires legacy versions.

Tool calling & vision

Native Bedrock support for tool calling (ToolConfiguration) and vision (ImageBlock) is not exposed on the IAIProvider surface — that contract is text-only. The agent loop in Compendium.Application owns tool calling; a typed Bedrock-native client surfacing vision and tools will land in a follow-up preview.

Compatibility

License

MIT — see LICENSE.