Using Custom Brains

​

Overview

Our of the box Anam provides an LLM to act as the brains of your personas. This is great for most use cases and helps produce the lowest possible latency when having conversations. However, you may already have your own LLM or other intelligence that you want to integrate with your persona. Anam allows you to do this by providing a way to disable the default brains and use the SDK’s talk() method to give speech instructions directly to the persona.

​

The Process

Using your own custom brains requires three steps to configure:

1. Disable the default Anam brains.
2. Listen for updates to the conversation history to know when you need to respond.
3. Use the talk method to instruct the persona to speak your custom response.

​

1. Disabling the Default Brain

You can turn off the Anam LLM by using the disableBrains configuration option when initialising the Anam client. When disabled, the persona will simply listen to user input and only respond to explicit talk commands.

import { createClient } from '@anam-ai/js-sdk';

const anamClient = createClient('your-session-token', {
  personaId: 'your-persona-id',
  disableBrains: true
});

​

2. Listening for Conversation History

In order to know when to respond, you’ll need to listen for updates to the conversation. You can do this by adding a listener for one of the following events.

anamClient.addListener(AnamEvent.MESSAGE_HISTORY_UPDATED, async (messageHistory: Message[]) => {
  if (messageHistory.length > 0) {
      const latestMessage = messageHistory[messageHistory.length - 1];
      // only respond to user messages
      if (latestMessage.role === 'user') {
        // get the response from your custom brain
        const response = await getCustomBrainResponse(messageHistory);
        // instruct the persona to speak the response
        anamClient.talk(response);
      }
    }
});

​

3. Using Talk Commands

When using your own brain implementation, you’ll need to control the persona’s responses using the talk method. The most basic method is to pass the full string of your response to a single talk call.

anamClient.talk('Content to say');

​

Streaming Responses

For reduced latency when using custom LLMs, you can stream responses in chunks using the TalkMessageStream:

const talkMessageStream = anamClient.createTalkMessageStream();
const chunks = ['He', 'l', 'lo', ', how are you?'];

for (const chunk of chunks) {
  if (talkMessageStream.isActive()) {
    talkMessageStream.streamMessageChunk(
      chunk,
      chunk === chunks[chunks.length - 1]
    );
  }
}

​

Important Considerations

1. One TalkMessageStream represents one conversation turn
2. A turn can end in two ways:
   • End of speech: When streamMessageChunk is called with endOfSpeech set to true or the endMessage() method is called.
   • User interruption: When the user speaks during streaming.
3. Once a turn ends, the TalkMessageStream object is closed and cannot be reused.
4. Create a new TalkMessageStream for each new turn.

​

Correlation IDs

For tracking purposes, you can attach a correlation ID to your talk streams by passing the optional correlationId parameter to the createTalkMessageStream method:

const talkMessageStream = anamClient.createTalkMessageStream("unique-id-123");

​

Handling Interrupts

When streaming talk commands your users may interrupt the persona and produce new speech events that you need to respond with. You can listen for interrupt events when implementing your own brain which can let you know when you no longer need to continue using the current talk stream.

anamClient.addListener(AnamEvent.TALK_STREAM_INTERRUPTED, (event) => {
  console.log('Stream interrupted, correlation ID:', event.correlationId);
});

​

Next Steps