Introduction

Helpers provide developer convenience to extend the functionality of the base RTVI Client by wrapping actions, messages or other arbitrary features. They define a common interface for working with transports, events, and messages within the context of a specific service. For example, you may have a LLM helper, TTS helper, and so on.

Since RTVI does not make any assumptions about which services a developer will leverage in their bots, helpers remove the need for unnecessary service-centric methods and handlers in the base client, and facilitate framework extensibility.

Some examples of where helpers can be useful include:

• LLM message context modificiation
• Direct text-to-speech output
• Multi-lingual workflows
• Custom convenience functions

It’s easiest to think of helpers as wrappers around existing client functionality that target a specific service, making your RTVI codebase less verbose.

Since most pipelines will feature at least one LLM, the core RTVI web package includes a single LLM helper out of the box. Refer to LLM Helper for a quick-start introduction on how to work with LLMs in your project.

Using a helper

Helpers must be imported, defined and registered to your client.

Here is an example using the bundled LLM helper:

import {
  LLMHelper,
  LLMHelperEvent
  RTVIMessage,
  RTVIClient,
} from "realtime-ai";

const rtviClient = new RTVIClient({
  // ...
});

const llmHelper = new LLMHelper({
  callbacks: {
      onLLMContextUpdate: (messages) => {
          // ...
      }
  }
});

rtviClient.registerHelper("llm", llmHelper);

// Events
rtviClient.on(LLMHelperEvent.LLMContextUpdate, (messages) => console.log(messages));

await llmHelper.appendContext({
  role: "user",
  content: "Tell me a joke!",
});

Registering a helper

Registering a helper creates a reference with your RTVI client. Helpers must be registered in order to use event and action dispatch / resolution.

rtviClient.registerHelper(service, helper);

service · string · required

Service targeted by the helper as part of your voice client constructor services map. An exception is thrown if no matching service can be found. Note: targeting services is required for scenarios where a pipeline contains multiple declarations of the same service type. For example, your pipeline may have two or more LLM services registered.

helper · RTVIClientHelper · required

A helper subclass instance of RTVIClientHelper.

Retrieving a helper

The client holds a reference to all registered helpers. This can be useful if you do not want to hold / pass a reference across your codebase:

const llmHelper = rtviClient.getHelper("llm") as LLMHelper;

Unregistering a helper

A service in your pipeline can currently have reference to a single helper to avoid conflicting action dispatch.

You can unregister a helper from a service like so:

rtviClient.unregisterHelper("llm");

Unregistering helpers will not delete any references in your own code. Any subsequent attempts to use an unregistered helper against the service name will raise an exception. If you register a new helper with the same service name and use a previous helper instance, you may encounter errors.

How are helper messages handled?

The RTVIClientHelper base class requires that a helper define a handleMessage method.

Any RTVI messages received from the transport that does not match native voice client message types are passed through to your client’s registered helpers:

public handleMessage(ev: VoiceMessage): void {
  switch(ev.type) {
    case MyHelperMessageType.SOME_MESSAGE:
      this._options.callbacks.onSomeMessage(ev.data);
      break;
      // etc
  }
}

Be mindful of defining a ‘default’ switch handler if you have multiple registered helpers listening for the same message type. Catching them will prevent lower order helpers from receiving the message.

---

Writing helpers

Creating your own helper classes can be useful when you want to create reusable code that targets a specific service (or series of services.)

The abstract RTVIClientHelper class is intentionally lightweight to provide as much flexibility to the developer as needed.

A typical helper definition will likely include:

• A constructor
• Callback methods
• Event and message types

A basic implementation of a helper would look like this:

myhelper.ts

import { 
  RTVIClientHelper,
  RTVIClientHelperOptions
} from "realtime-ai";

// --- Events
// E.g. rtviClient.on(MyHelperEvents.SomeEvent, (message) => { ... });
export enum MyHelperEvents {
SomeEvent = "some-event",
}

// --- Message types
// E.g. await rtviClient.action({ service: "myService", action: MyHelperMessageType.SOME_MESSAGE, arguments: [...]});
export enum MyHelperMessageType {
SOME_MESSAGE = "some-message",
}

// --- Callbacks
export type MyHelperCallbacks = Partial<{
onSomeMessage: (message) => void;
}>;

// --- Interface and class
export interface MyHelperOptions extends RTVIClientHelperOptions {
callbacks?: MyHelperCallbacks;
}

export class LLMHelper extends RTVIClientHelper {
  protected declare _options: MyHelperOptions;

constructor(options: MyHelperOptions) {
super(options);
}

public handleMessage(ev: VoiceMessage): void {
switch (ev.type) {
case MyHelperMessageType.SOME_MESSAGE:
// dispatch action response etc
break;
}
}

public getMessageTypes(): string[] {
return Object.values(MyHelperMessageType) as string[];
}
}