Documentation / Conversation flow
Conversation flow
Adjust silence, turn-taking, interruptions and conversation duration.
The prompt defines what to say. Conversation settings determine when to listen, respond and stop because of a time limit.
Before you start
Open the agent settings and run a voice test in a quiet environment. Save changes before comparing results.
Choose the right control
| Control | What it changes |
|---|---|
| Maximum duration | Total call duration, not the time allowed for one reply |
| User inactivity | How long the user may remain silent |
| Turn start | Strategy for recognizing that the user started speaking |
| Turn stop | Strategy for deciding that the user finished speaking |
| Interruption | Whether user speech can interrupt the agent; also check the step's option |
The agent configuration's max_call_duration defaults to 300 seconds, with a ceiling of 1,200 seconds. Usage limits authorized for a run may reduce its available duration.
max_user_idle_timeout defaults to 10 seconds. The backend defines default, min_words and provisional_vad turn-start strategies, and transcription and turn_analyzer turn-stop strategies. Choose from the options offered by your configuration.
Tune using a test conversation
- Ask a short question and let the answer finish.
- Say a sentence with a pause in the middle. Check whether the agent replies before you finish.
- Interrupt a long answer and verify that the step permits interruption.
- Stay silent and observe the inactivity limit.
- Change one control and repeat the same scenario.
Premature replies call for reviewing turn-stop behavior. Delay after a complete sentence calls for reviewing timing and model/tool latency. Avoid trying to fix both only through the prompt.
Text chat
Chat does not use speech detection. It has its own inactivity setting, text_chat_inactivity_timeout_seconds. Audio timing does not apply to this mode.
After testing, publish the configuration. Use Troubleshooting if behavior differs from your expectations.