ZenTalk Docs
Agents

Advanced Settings

Fine-tune model parameters, voice activity detection, and call behavior.

The Advanced tab provides granular control over your agent's behavior.

Model Settings

Control the LLM's response generation:

SettingRangeDefaultDescription
Temperature0–20.7Higher = more creative, lower = more focused
Max Tokens1–81921024Maximum response length
Frequency Penalty-2 to 20Reduces word repetition
Presence Penalty-2 to 20Encourages topic diversity
Extended ThinkingOn/OffOffEnables chain-of-thought reasoning
Reasoning EffortLow/Medium/HighMediumControls thinking depth (when enabled)

Voice Activity Detection (VAD)

Controls how the system detects when the user starts and stops speaking:

SettingRangeDefaultDescription
STT Latency0–3s0Delay before processing speech
VAD Confidence0–10.7Minimum confidence to detect speech
VAD Start Duration0.1–1.0s0.2How long speech must last to be detected
VAD Stop Duration0.1–2.0s0.2Silence duration before end-of-speech
VAD Min Volume0–10.6Minimum audio level to detect speech
Smart Turn Stop0.5–5.0s1.5Wait time before the agent responds

Tuning Tips

  • Noisy environments: Increase VAD Confidence and Min Volume
  • Slow speakers: Increase VAD Stop Duration and Smart Turn Stop
  • Fast conversations: Decrease Smart Turn Stop for quicker responses

Conversation Behavior

SettingDefaultDescription
Allow InterruptionsOnLet users interrupt the agent mid-sentence
Mute During WelcomeOnPrevent interruption during the welcome message
Mute During Function CallsOnSilence while tools execute
TTS Pace1.0xSpeech speed (0.5x–2.0x)

Call Management

SettingDefaultDescription
Idle Timeout300sEnd call after this much silence
Cancel on IdleOnAuto-hang-up when idle timeout is reached
User Idle DetectionOnDetect when the user goes quiet
User Idle Timeout10sSeconds of user silence before prompting
User Idle Retries3How many times to prompt before ending
Max Call Duration0 (unlimited)Hard limit on call length

Data & Recording

SettingDefaultDescription
Record CallsOffSave stereo recordings (user + bot) to MinIO/S3
Save TranscriptsOnStore conversation transcripts in the database
Save MetricsOnTrack latency, tokens, and performance data
Context SummarizationOffSummarize long conversations to save tokens
Max Unsummarized Messages20Messages before auto-summarization kicks in