​

What are function calls?

Function calls allow modern Large Language Models (LLMs) like ChatGPT to perform tasks outside of text generation. For example, if I create a math assistant with ChatGPT, I might make a “multiply” function that ChatGPT can call to multiply two numbers together.

If you add an action to a vocode agent, its corresponding function calling schema is added to each ChatGPT query. If the function call is outputted by ChatGPT, the corresponding action is triggered automatically.

Note: In vocode agents, ChatGPT can return a text response alongside a function call. In this scenario, the text is synthesized and played first, then the action is run.

​

What are phrase triggers?

Phrase triggers are text phrases that activate an action when produced by an agent. Phrase triggers are useful in production use-cases where consistency and reliability are important, since text outputs are easier to control in comparison to function calls.

​

Configuring action triggers

Each vocode action config contains an action_trigger field to specify how the action is triggered. The default trigger is function calling. The examples below demonstrate how to set action triggers for an EndConversation action config:

Function calls:

from vocode.streaming.action.end_conversation import EndConversationVocodeActionConfig
from vocode.streaming.models.actions import FunctionCallActionTrigger

EndConversationVocodeActionConfig(
    type="action_end_conversation",
    action_trigger=FunctionCallActionTrigger(
        type="action_trigger_function_call"
    )
)

Note: You can also leave action_trigger field empty and vocode will default to function calls.

Phrase triggers:

from vocode.streaming.action.end_conversation import EndConversationVocodeActionConfig
from vocode.streaming.models.actions import PhraseBasedActionTrigger, PhraseBasedActionTriggerConfig, PhraseTrigger

EndConversationVocodeActionConfig(
    type="action_end_conversation",
    action_trigger=PhraseBasedActionTrigger(
        type = "action_trigger_phrase_based",
        config: PhraseBasedActionTriggerConfig(
            phrase_triggers = [
                PhraseTrigger(
                    phrase="Ending conversation now",
                    condition="phrase_condition_type_contains"
                ),
                # Additional phrase triggers can be listed here
            ]
        )
    )
)

For the code above, if the agent says ‘Ending conversation now’, the end conversation action will automatically be taken. You can add multiple phrase triggers for an action by passing a list of PhraseTrigger instances.

The phrase_condition_type_contains condition configures the agent to run the action if its output contains the phrase. So, the action will also be run if the agent says ‘I am ending conversation now’.

​

Example Scenario

Let’s assume we have an action called TurnOnLight that turns on your bedroom light when activated. Take the following conversation:

Human: Hello
AI: How can I assist you today?
Human: Please turn on the lights.

If the action is triggered via a function call, the bot may respond with:

AI: Sure!
BOT_ACTION_START: Running `action_turn_on_light`

or it may run the action without saying anything.

If we use a phrase trigger instead, we may do the following:

1. In our prompt, write “If the human asks to turn on the lights, say ‘I will turn on the lights now’ verbatim”
2. Configure TurnOnLight with “I will turn on the lights now” as its phrase trigger

Then, the bot will respond with:

AI: I will turn on the lights now.
BOT_ACTION_START: Running `action_turn_on_light`