AWS.Bedrock.MEAI 1.1.2

AWS.Bedrock.MEAI

AWS.Bedrock.MEAI provides Microsoft.Extensions.AI implementations for Amazon Bedrock. It lets .NET developers use Bedrock models through Microsoft's standard AI abstractions — IChatClient, IEmbeddingGenerator, IImageGenerator, and IRealtimeClient — so Bedrock plugs into the broader .NET AI ecosystem (agents, middleware, tooling) with no Bedrock-specific glue code.

This package was previously published as AWSSDK.Extensions.Bedrock.MEAI from the aws-sdk-net repository. It now lives in aws-dotnet-ai alongside the other .NET AI packages and releases independently. The AmazonBedrockRuntimeExtensions methods (AsIChatClient, AsIEmbeddingGenerator, etc.) remain in the Amazon.BedrockRuntime namespace, so existing call sites need no code change — just update the package reference. The package's other types (such as BedrockStructuredOutputMode) live in the AWS.Bedrock.MEAI namespace; add using AWS.Bedrock.MEAI; only where you reference those directly.

Installation

dotnet add package AWS.Bedrock.MEAI

Getting started

Turn an IAmazonBedrockRuntime client into an IChatClient:

using Amazon.BedrockRuntime;
using Microsoft.Extensions.AI;

IAmazonBedrockRuntime runtime = new AmazonBedrockRuntimeClient();

IChatClient chatClient = runtime.AsIChatClient("anthropic.claude-3-5-sonnet-20241022-v2:0");

var response = await chatClient.GetResponseAsync("What is Amazon Bedrock?");
Console.WriteLine(response);

Generate embeddings:

IEmbeddingGenerator<string, Embedding<float>> generator =
    runtime.AsIEmbeddingGenerator("amazon.titan-embed-text-v2:0");

var embeddings = await generator.GenerateAsync(["Hello world"]);

Features

• Chat (AsIChatClient) over the Bedrock Converse / ConverseStream APIs, including tool/function calling, multi-modal content (image, video, document), reasoning / extended thinking, citations, and prompt cache points.
• Structured output via JSON Schema, in two modes (BedrockStructuredOutputMode):
  • SyntheticTool (default) — works across the broadest set of models.
  • Native — uses Bedrock native structured outputs on models that support it; composes with user-provided tools and supports streaming.
• Message coalescing (opt-in) — pass a BedrockChatClientOptions with CoalesceConsecutiveMessages = true to AsIChatClient to combine consecutive request messages that map to the same Converse role (for example a tool-result message followed by a user message, both of which map to the Converse user role) into a single message, giving replayed histories a normalized request shape. Content-block order is preserved, messages separated by a cache point are never combined, and caller-supplied raw messages are not modified. Defaults to false (each message is sent separately). Optional client settings live on BedrockChatClientOptions so future options can be added without new AsIChatClient overloads.
• Embeddings (AsIEmbeddingGenerator).
• Native stop reason and guardrail trace surfaced on ChatResponse.AdditionalProperties (and the matching streaming updates) under the AmazonBedrockRuntimeExtensions.StopReasonKey ("StopReason") and TraceKey ("Trace") keys. Because multiple native stop reasons map to one ChatFinishReason (both guardrail_intervened and content_filtered map to ContentFilter), the StopReason string lets you tell them apart; Trace carries the guardrail trace as a JsonElement serialized from the SDK trace object, so it serializes safely when the response is persisted (including under trimming / Native AOT).
• Image generation (AsIImageGenerator) for Stability and Amazon Titan / Nova Canvas models. (Experimental — MEAI001.)
• Realtime audio (AsIRealtimeClient) for Amazon Nova Sonic bidirectional streaming, on net8.0 and later. (Experimental — MEAI001.)

Some APIs are annotated [Experimental("MEAI001")] because they build on evaluation-stage Microsoft.Extensions.AI abstractions. Suppress MEAI001 to use them.

Supported target frameworks

net472, netstandard2.0, and net8.0. The realtime (IRealtimeClient) APIs require net8.0 or later.

Getting help

Open an issue on the aws-dotnet-ai repository.

License

Apache-2.0. See LICENSE.