Universal.Operative.Sdk.Realtime.AwsBedrock 1.0.0

About

Universal.Operative.Sdk.Realtime.AwsBedrock adapts the Amazon Bedrock Nova Sonic bidirectional streaming API to Universal.Operative.Sdk.Realtime's provider-neutral IModel/IRealtimeSession/SessionEvent model.

Unlike the OpenAI and Gemini adapters, Bedrock's InvokeModelWithBidirectionalStreamAsync is a generic byte-stream API with no Nova-Sonic-specific vendor types -- every event (sessionStart, promptStart, contentStart, audioInput, toolResult, ...) is JSON this adapter constructs and parses directly, grounded in AWS's own input/output event documentation.

How to Use

Installation

dotnet add package Universal.Operative.Sdk
dotnet add package Universal.Operative.Sdk.Realtime
dotnet add package Universal.Operative.Sdk.Realtime.AwsBedrock

Quick start

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

// Bare model id -- Nova Sonic has neither a geo nor a global cross-region
// inference profile (per AWS's own model card), so a "us."/"eu." prefix is
// rejected, not just unnecessary. Region must be one Nova Sonic actually runs
// in: us-east-1, eu-north-1, or ap-northeast-1 as of this writing.
var model = new Model(new Model.Options
{
    ModelId = "amazon.nova-sonic-v1:0",
    Region = "us-east-1",
});

IRealtimeSession session = await model.ConnectAsync(new RealtimeSessionOptions
{
    SystemPrompt = [new TextBlock("You are a helpful voice assistant.")],
    Tools = [myTool],
});

await foreach (SessionOutputEvent e in session.Output)
{
    switch (e)
    {
        case SessionReadyEvent: Console.WriteLine("connected"); break;
        case TranscriptDeltaEvent t: Console.WriteLine($"{t.Role}: {t.Text}"); break;
        case AudioOutputDeltaEvent a: /* play a.Audio (24kHz PCM16 by default) */ break;
        case InterruptedEvent: /* stop playback immediately */ break;
    }
}

await session.SendAsync(new AppendAudioInputEvent { Audio = micChunk });
await session.SendAsync(new CommitAudioInputEvent());

Fidelity notes

• Capabilities: AudioInProcess | ToolCalling | Transcripts -- no resumption capability flags are set. Nova Sonic's connections are capped at 8 minutes; this adapter raises a synthetic SessionExpiringEvent as that approaches, but does not (yet) implement the client-assisted splice/replay AWS's own samples use to hop to a fresh connection -- a caller must reconnect and replay its own history.
• Interruption: Nova Sonic documents barge-in as a contentEnd.stopReason == "INTERRUPTED" on the in-flight text content block -- mapped to a Confirmed-confidence InterruptedEvent, not a heuristic.
• Tool results: the full ToolResult content is flattened into a single JSON object ({"result": "..."} or {"error": "..."}) sent as Nova Sonic's toolResult event content, mirroring the Gemini adapter.
• Raw passthrough: usageEvent, per-content-block contentStart/contentEnd events that aren't the interruption signal, and anything else not explicitly mapped arrive as a RawPassthroughEvent rather than being dropped.
• Tool choice: RealtimeSessionOptions.ToolChoice throws NotSupportedException -- Nova Sonic's promptStart.toolConfiguration only ever carries a tools array; no mode/choice field exists anywhere in its schema.
• Turn completion: TurnCompletedEvent is raised from contentEnd.stopReason == "END_TURN", not from completionEnd alone -- completionEnd doesn't always arrive even after a full response has been delivered.
• Continuous audio input: Nova Sonic closes the connection if no audioInput event arrives for about a minute while a prompt is open. While an audio content block is open, Session automatically sends a short silence chunk during any gap to keep the connection alive; it does not open a block on its own.
• Cross-modal text: sending text while no audio content block is open gets no response at all, matching AWS's documentation that cross-modal input requires an active streaming session. Even with an audio block open, a response to cross-modal text is occasionally absent.