Skip to main content

LLM in telephony

Beta

In the phone channel, you can use the  llmRequest reply type so that the bot receives text from the LLM and synthesizes speech in streaming mode.

To get started:

  1. Add a secret for Caila.
  2. Use the llmRequest reply type in your script.

LLM and TTS providers​

  • To generate texts, you can currently only access models of the openai-proxy service on the Caila platform.

    caution

    LLMs are only available in the paid plan. To use the models, top up your balance in Caila.

  • Speech synthesis works with any TTS provider.

llmRequest advantages​

The llmRequest reply type generates text and synthesizes speech in streaming mode. The bot receives text from the LLM by sentences and synthesizes speech by sentences as well. The two processes occur in parallel.

This way you can reduce pauses before bot responses compared to when the generation and synthesis are done sequentially.

Example of sequential generation and synthesis without llmRequest
// Bot receives text from LLM
var llmResponse = $gpt.createChatCompletion([{ "role": "user", "content": $request.query }]);
var response = llmResponse.choices[0].message.content;
// Bot synthesizes speech for entire text
$reactions.answer(response);

Here, the bot:

  1. Accesses the LLM via $gpt.createChatCompletion. The bot is waiting for the LLM to generate the text completely.
  2. Sends the entire text to speech synthesis and waits until speech is synthesized for the entire text.
  3. Plays the text.

In this case, there might be a long pause of several seconds between the user request and the bot response.

llmRequest also lets you specify phrases that the bot will say to fill the pause at the beginning of text generation.

Access token for Caila​

To use services and generative models from Caila in third-party applications, including JAICP, you need a personal access token. To issue a token:

  1. Go to Caila.

    tip

    Caila and Conversational Cloud use a shared account base, so if you are registered on Conversational Cloud, you do not need to register additionally on Caila.

  2. Go to the My space → API tokens section.

  3. In the upper right corner, click Create token.

  4. Give a name to the token, generate it, and copy it to the clipboard.

Next, add this token to JAICP:

  1. Go to JAICP. In the Secrets and variables section add a new secret.

  2. In the llmRequest reply, specify the secret name in the tokenSecret property.

Use llmRequest in script​

state: NoMatch
    event!: noMatch
    script:
        $response.replies = $response.replies || [];
        $response.replies.push({
            type: "llmRequest",
            provider: "CAILA_OPEN_AI",
            // Text generation model
            model: "gpt-4o",
            // Secret name
            tokenSecret: "MY_LLM_TOKEN",
            // Prompt and user request
            messages: [
                {"role": "system", "content": "Keep answers short. A few sentences at most."},
                {"role": "user","content": $request.query}
            ]
        });

In this example, llmRequest is used in the NoMatch state:

  1. The bot sends a request to generate text to the openai-proxy service on the Caila platform. The messages field contains:

    • Prompt for the LLM to make the model generate a short answer.
    • User request text stored in $request.query.

  2. Once the bot receives the first sentence from LLM, the bot will start synthesizing speech.

  3. The bot plays the first sentence to the user.

  4. The bot continues to synthesize and play speech by sentences until it receives the entire text from LLM.

caution

After transitioning to the state, the bot immediately starts preparing text and speech for llmRequest. The Caila and speech synthesis limits can be charged even if the user ended the call and the bot did not play the speech.

Fill pauses​

In the script, a pause might occur while the bot is waiting for the first sentence of text from the LLM. There are two ways to fill this pause:

  • Use the fillersPhraseConfig setting. You can specify a phrase that the bot will say at the beginning of generation. This will help fill the pause if it is too long.

    state: NoMatch
        event!: noMatch
        script:
            $response.replies = $response.replies || [];
            $response.replies.push({
                type: "llmRequest",
                …
                // The bot says the phrase if the pause exceeds 2000 ms.
                fillersPhraseConfig: {"fillerPhrase": "Great question!", "activationDelayMs": 2000}
            });

  • Specify other responses before llmRequest. After transitioning to the state, the bot immediately starts preparing text and speech for llmRequest. The bot can perform other reactions before llmRequest while waiting for a response from the LLM.

    state: NoMatch
        event!: noMatch
        # The bot stars generating the llmRequest response right after transitioning to the state
        a: Great question!
        a: Let me think
        # The bot has already said 2 phrases. During this time, it has prepared some part of the answer
        script:
            $response.replies = $response.replies || [];
            $response.replies.push({
                type: "llmRequest",
                …
            });

Barge-in​

A user can interrupt the bot if the bot plays speech via llmRequest.

In the $dialer.bargeInResponse method, set the forced barge-in mode. If the user interrupts the bot, the bot stops the speech and does not play the LLM response to the end:

state: NoMatch
    event!: noMatch
    script:
        // Barge-in settings
        $dialer.bargeInResponse({
            bargeIn: "forced",
            bargeInTrigger: "final",
            noInterruptTime: 0
        });
        // llmRequest response
        $response.replies = $response.replies || [];
        $response.replies.push({
            type: "llmRequest",
            provider: "CAILA_OPEN_AI",
            model: "gpt-4o",
            tokenSecret: "MY_LLM_TOKEN",
            messages: [{"role": "user", "content": $request.query}],
        });

Conditional barge-in​

You can also configure conditional barge-in. To do this, pass a bargeInReply object to llmRequest.

In the example below:

  1. The bot creates an empty response with the bargeInIf parameter.
  2. The bot extracts a bargeInReply object from this response and passes the object to llmRequest.
  3. The barge-in is activated if the user request contains “support agent”.
state: NoMatch
    event!: noMatch
    # Create an empty response with the bargeInIf parameter
    a: || bargeInIf = "LLM response"
    script:
        // Barge-in settings
        $dialer.bargeInResponse({
            bargeIn: "forced",
            bargeInTrigger: "final",
            noInterruptTime: 0
        });
        // Save bargeInReply from the empty response and delete the empty response
        var bargeInReplyObject = $response.replies.pop().bargeInReply;
        // llmRequest response
        $response.replies = $response.replies || [];
        $response.replies.push({
            type: "llmRequest",
            provider: "CAILA_OPEN_AI",
            model: "gpt-4o",
            tokenSecret: "MY_LLM_TOKEN",
            messages: [
                {"role": "user","content": $request.query}
            ],
            // Pass the bargeInReply object
            bargeInReply: bargeInReplyObject
        });

state: BargeInCondition || noContext = true
    event!: bargeInIntent
    script:
        var text = $dialer.getBargeInIntentStatus().text;
        // The barge-in is activated if the user request contains “support agent”
        if (text.indexOf("support agent") > -1) {
            $dialer.bargeInInterrupt(true);
        }
Restrictions