RATT Agent Library

Welcome to the **RATT Agent Library **! 🎧⚡ This tiny TypeScript client streams microphone (or external) audio over WebSocket, handles reconnection and heartbeats, and emits typed events for live transcription and UI state. It works in browsers (with AudioWorklets) and in Node (via external PCM).

🚀 Features

• Low-latency audio streaming (16 kHz, 16-bit PCM) over WebSocket
• Browser + Node: use the mic in browsers or push external PCM in any runtime
• Typed events: ready, mic state, amplitude (VAD energy), transcription, socket messages, errors
• Prebuffering & gating: capture early, start sending when the server says start_audio
• Resilient WS: single-flight connect, StrictMode-safe reuse, heartbeats & auto-reconnect
• Safe defaults: echo cancellation, AGC, noise suppression (browser)
• Tiny API: a single class AssistantClient you can drop into your app

📦 Installation

npm i ratt-lib
# or
pnpm add ratt-lib

✨ Quick Start (Browser)

import { AssistantClient, AssistantOptions, AssistantEvent } from "ratt-lib";

// Simple mutable ref for requestId (works well with React useRef) and this id will be unique for each req
const requestId = { current: "" };

const chatSessionId = "test"; //Unique for whole session
const clientId = "test"; //Unique for whole session

// ✅ Build the REQUIRED rattAgentDetails object
const rattAgentDetails = {
    conciergeId: chatbotData?.id, // Your chatbot Id
    conciergeName: chatbotData?.name ?? chatbotData?.assistantDetails?.name,
    organizationId: chatbotData?.organization,
    organizationName: chatbotData?.organizationName,
    requestId: requestId.current, // will be overridden with a fresh value when session starts
    agentSettings: {
        voiceAgentMongoId: chatbotData?.agents?.filter((a: any) => a.title === "RATTAgent")?.[0]?._id,
    },
    username: user?.provider?.name, // user name
    useremailId: user?.email, // user email
    chatSessionId: chatSessionId,
    rlefVoiceTaskId: chatbotData?.audioTranscription?.modelId || CREATE_AUDIO_RELF_VOICE_TASK_ID_DEFAULT_VALUE, // rlef model id
    assistant_type: chatbotData?.assistant_type, // your assistant type
    isAudioRequest: true, //always true
    client_id: clientId,
    userId: encodeParam(USER_ID_DEFAULT_VALUE), // your user id
    // keep below three as it is if test instant learning is not there , otherwise please send these values as well as per the requirements.
    testQuestion: "",
    testAnswer: "",
    testVariants: JSON.stringify({ Edit: [], Add: [], Delete: [] }),
} as const;

const client = new AssistantClient({
    url: "wss://dev-egpt.techo.camp/audioStreamingWebsocket?clientId=${clientId}&sessionId=${chatSessionId}", // Required: your WS endpoint
    requestId, // Required: { current: string }
    rattAgentDetails, // ✅ pass the full required object
    onSend: () => console.log("Transcript submitted"),
    showToast: (type, title, msg) => console.log(type, title, msg),
    // optional tunables:
    pingIntervalMs: 5000,
    maxMissedPongs: 2,
    workletBasePath: "/", // where recorder/vad worklets are served from
});

// Listen to events
client.on(AssistantEvent.READY, () => {
    console.log("WS ready:", client.wsReady);
});

client.on(AssistantEvent.MIC_CONNECTING, ({ detail }) => {
    console.log("Mic connecting:", detail.connecting);
});

client.on(AssistantEvent.MIC_OPEN, ({ detail }) => {
    console.log("Mic open:", detail.open);
});

client.on(AssistantEvent.AMPLITUDE, ({ detail }) => {
    // 0..~1 energy (not dB) to drive a mic meter UI
    console.log("Amplitude:", detail.value);
});

client.on(AssistantEvent.TRANSCRIPTION, ({ detail }) => {
    // progressive or final text
    console.log("Transcript:", detail.text);
});

client.on(AssistantEvent.ERROR, ({ detail }) => {
    console.error("Assistant error:", detail.error);
});

// Start a session (ask server to begin, then it will reply with start_audio)
document.querySelector("#start")!.addEventListener("click", () => {
    client.startSession(); // toggles ON (opens mic flow) basically handleMicClick function
});

// Stop (ask server to disconnect and teardown locally)
document.querySelector("#stop")!.addEventListener("click", () => {
    client.stopAudio(); // toggles OFF
});

What happens under the hood?

1. startSession() sends your rattAgentDetails + a new requestId.

2. When your server responds with { "start_audio": true }, the client:

   • marks the mic as open,
   • starts converting to PCM16,
   • and begins streaming.

3. As your server streams partial ASR, send either:

   • {"streaming_data":{"previous_transcription":"...", "new_transcription":"..."}} (chunked delta) or
   • {"transcription":"final text"} (full updates)

4. End with {"stop_audio": true} and/or {"disconnect": true} when you’re done.

🧩 React one-liner (optional)

useEffect(() => {
    const unsub = client.on(AssistantEvent.TRANSCRIPTION, ({ detail }) => {
        setText(detail.text);
    });
    return unsub; // cleanly removes listener
}, []);

🖥️ Node / External Audio (no browser mic)

If you're not in a browser (or you have your own capture pipeline), set externalAudio: true and push PCM yourself.

import { AssistantClient } from "ratt-lib";
import fs from "node:fs";

const requestId = { current: "" };
const client = new AssistantClient({
    url: "wss://dev-egpt.techo.camp/audioStreamingWebsocket?clientId=${clientId}&sessionId=${chatSessionId}",
    requestId,
    externalAudio: true, // ⬅️ no AudioContext, no mic
    externalAmplitudeRms: true, // optional: compute amplitude from PCM
});

// Start session -> wait for server {"start_audio": true}
await client.connect();
await client.startSession();

// Now stream your PCM16 mono 16kHz data
const pcm = fs.readFileSync("./audio.raw"); // Int16 little-endian mono 16kHz
client.pushPCM16(new Int16Array(pcm.buffer, pcm.byteOffset, pcm.byteLength / 2));

// ...push more chunks as they arrive...
// When done: or it will be auto stop by server when it detects some silence is there
await client.stopAudio();

PCM format: 16 kHz, 16-bit, mono, little-endian. If you have Float32Array [-1..1], call pushFloat32() instead.

🔔 Events

All events are emitted as standard CustomEvents and strongly typed via AssistantEvents type.

• READY — WebSocket is ready (connected & open)
• MIC_CONNECTING — { connecting: boolean } while we prep/prompt for mic
• MIC_OPEN — { open: boolean } mic flow is active/inactive
• AMPLITUDE — { value: number } live energy (for a mic meter)
• TRANSCRIPTION — { text: string, delta?: string } progressive or final
• SOCKET_MESSAGE — { raw: MessageEvent, parsed?: any } every incoming WS message
• ERROR — { error: unknown } any operational error

const off = client.on(AssistantEvent.TRANSCRIPTION, ({ detail }) => {
    console.log(detail.text);
});
off(); // unsubscribe

⚙️ Options

type AssistantOptions = {
    url: string; // WS endpoint
    requestId: { current: string }; // mutable ref; client writes new ID per session
    rattAgentDetails?: Record<string, any>;
    onSend?: () => void; // called when server requests "disconnect"
    showToast?: (type: "error" | "info" | "success", title: string, msg: string) => void;

    // Connection / heartbeat
    pingIntervalMs?: number; // default 5000
    maxMissedPongs?: number; // default 2

    // Audio (browser)
    workletBasePath?: string; // default "/"
    mediaStreamProvider?: () => Promise<MediaStream>; // default: getUserMedia (1ch, 16k, AGC/NS/EC enabled)
    audioContextFactory?: () => AudioContext | null; // default: new AudioContext() in browser, null in Node
    workletLoader?: (base: string) => Promise<AudioContext | null>; // default: ensureAudioContextAndWorklets

    // External audio (Node or custom capture)
    externalAudio?: boolean; // default: false in browser, true in Node
    externalAmplitudeRms?: boolean; // default: true
    pcmChunkSize?: number; // default: TARGET_SAMPLES (typically 16000)
};

🧪 Common Methods

await client.connect(); // single-flight; reuses active WS if present
await client.startSession(); // begin mic flow or external audio session
await client.stopAudio(); // stop current session (sends {disconnect:true})
client.disconnect(); // local teardown (no forced WS close)
client.teardown(); // local teardown helpers
client.closeSocket(); // forcibly close WS and detach handlers

// Mic helpers
await client.beginPrebuffering(); // start capturing locally; don't send yet
client.stopPrebuffering(); // stop & clear buffered audio
await client.startMic(); // explicitly start mic capture
client.stopMic(); // explicitly stop mic (and send {disconnect:true})

// External audio
client.pushPCM16(int16ArrayOrBuffer);
client.pushFloat32(float32Array);

// State getters
client.wsReady; // boolean
client.micOpen; // boolean
client.micConnecting; // boolean
client.amplitude; // number (0..~1)
client.transcription; // latest accumulated text

🧯 Troubleshooting

• No audio sent Ensure your server replies with {"start_audio": true}. The client buffers until the gate opens.
• Mic blocked Browser will throw NotAllowedError. The client emits ERROR and calls showToast(...).
• Noisy audio / echo The default constraints enable echo cancellation, AGC, and noise suppression. Override mediaStreamProvider if needed.
• Multiple connects in React StrictMode This client reuses a global ACTIVE_WS and a single CONNECT_PROMISE—you’re safe.
• Heartbeat timeouts Increase pingIntervalMs or maxMissedPongs if your WS hops are choppy.

🔐 Permissions & Security

• Browsers require a user gesture to start the microphone.
• Only minimal audio data is sent; handle it securely on your server. Use wss:// in production.

👥 Community & Support

Questions, bugs, or ideas? Open an issue in your repo or ping your team chat. Happy streaming! 🎙️💬