Build your agent

Interruptions

Choose, per speaking step, whether the caller can cut the agent off, and when.

Every step that speaks (Say, AI reply) has its own interruption_type, so a disclaimer can be protected while normal chat stays interruptible. An AI reply can also choose it per reply by returning interruption_type.

Types

interruption_type	Behaviour
`full`	Stop and listen. The agent stops as soon as the caller starts talking.
`no`	Always finish. The agent never stops mid-reply. Good for legal lines and key details.
`duration`	Only at the start. The caller can cut in during the first few seconds; after that the agent finishes.
`inverse_duration`	Only after a few seconds. The opening of every reply is protected; after that the caller can cut in.
`word_length`	Only on short replies. Short replies can be interrupted; longer ones are always finished.
`inverse_word_length`	Only on long replies. Short replies are always finished; the caller can cut into longer ones.

The numbers

durationnumber (seconds): For duration and inverse_duration: the cut-off, counted from when the agent starts speaking.
lengthnumber (words): For word_length and inverse_word_length: replies shorter (or longer) than this many words can be interrupted.
ignore-wordsstring[]: Words that never count as interrupting, such as hmm, okay, haan. A space-separated string works too.

Once the agent has finished speaking, the caller can always talk; these rules only apply while it's speaking.

Example

Protect the first three seconds of a reply, and ignore back-channel words:

1"parameters": {
2  "out_dict": { "speak": "This call may be recorded for quality and training." },
3  "interruption_type": "inverse_duration",
4  "interruption_metadata": { "duration": 3, "ignore-words": ["hmm", "okay", "haan"] }
5}

Next steps

Turn-taking: reply timing.

Interruptions

#Types

#The numbers

#Example

#Next steps

Types

The numbers

Example

Next steps