1. General settings

LLM Settings

For now, you have a choice between five models from the OpenAI suite. Depending on the complexity of your agent, it will be more judicious to use one over the others.

Some guidelines:

• GPT-4o mini and GPT-3.5 Turbo will work very well for very simple use cases—interactions of less than 30 seconds—that do not require multi-prompts or an extensive prompt.Example: A temp agency agent tasked with calling temporary workers to check their availability.

• GPT-4o is the default model we recommend. Its latency-reliability ratio is excellent and it applies to all use cases.

• GPT-4 Turbo (legacy): We recommend GPT-4 Turbo for tasks that require processing a very large context. For example, we still use it to interpret a large amount of raw data that would be returned as context to an agent. Example: An appointment scheduling agent that queries the availability of a calendar returning thousands of time slots.

The improvement in GPT-4o performance and the increasing accessibility of o1 models will soon render GPT-4 obsolete.

Coming soon: Integration of GPT-4o mini real-time.

Temperature: Temperature it a parameter that controls how “deterministic” or “creative” the model’s output is when generating text. Lower temperatures (close to 0) make the output more deterministic and focused.

Voice Settings

Only French and English voices are supported for now. Contact us at [e-mail] to request support for other languages.

Important elements to consider when choosing voices:

• Reliability: While voice models are improving rapidly, they are still very imperfect in French. For example, intonations are much less natural than in English. Other issues include unexpected switching to English pronunciation for words written in French and poor pronunciation of numbers and dates.

• Latency: The better the voice quality, the slower the speech synthesis.

• Stability: The tone between two sentences can change completely. These strong variations are not acceptable in most production use cases.

• Cost: There is a 7x difference between the two providers we have currently selected:

We have tested a plethora of different providers. In French, we have currently selected two:

• Azure: Azure voices are by far the most stable and generally induce the least latency. Choose Yvette for a medical reception agent or appointment scheduling, for example.

  Coming soon: Integration of other neural voices

• ElevenLabs: ElevenLabs voices are among the most natural on the market. The multilingual models are also the most advanced. For French, we monitor the regular outputs and select the most convincing ones.

  If you want us to integrate new voices you have discovered, make a request on the Discord server.

Specific settings for ElevenLabs:

• Stability slider determines how stable the voice is and the randomness between each generation. Lowering this slider introduces a broader emotional range for the voice.

• Similarity Boost: The similarity slider dictates how closely the AI should adhere to the original voice when attempting to replicate it.

• Fast Option: We currently do not recommend using fast voices as they are less reliable and less stable. However, they may be necessary for certain use cases.

• Other settings:

  • Initial Message: The first message never requires generation. What first phrase do you want your agent to say when it answers/the caller answers.

  • Interruptible (beta): When enabled, your agent will stop speaking when interrupted. We do not recommend this setting for now and advise against it for telephone agents.

  • Agent Base Prompt: The base prompt is the persistent prompt throughout the conversation. If you are using multi-prompt, regardless of the task you are in, the base prompt will apply. In single prompt, the base prompt is not required. We use it internally to define the general behavior of the agent.[Example of Base prompt in single prompt]

  • Webhook URL:

    The Webhook URL is a critical integration point that allows your system to send real-time events to users or other services. By configuring a webhook URL, you enable your application to notify external systems about various events.

    Setting Up the Webhook URL:

    1. Provide the URL: Enter the endpoint URL where you want to receive the events.
    2. Configure Event Types: Specify which events you want to subscribe to.
    3. Authentication: If required, set up authentication to secure the webhook.
    4. Testing: Ensure that your endpoint correctly receives and processes the events.

    Example Usage:

    Suppose you have an external CRM system that needs to log all incoming calls. By setting up a webhook URL pointing to your CRM’s API endpoint, every time a phone_call_connected event occurs, the CRM can automatically log the call details without manual intervention.

    Ensure that your webhook endpoint is secure and can handle the expected load of incoming events to prevent data loss or service interruptions.