​
Configure Your Assistant

Use the assistant editor to choose the language pipeline, define call behavior, and control what is retained after a call. The API uses snake_case field names; dashboard selections are saved in the same shape.

​
Providers and credentials

Each assistant has an LLM (model), text-to-speech (voice), and speech-to-text (transcriber) configuration. Use a provider’s model and voice IDs exactly as supplied by the dashboard or provider. custom LLM, STT, and TTS configurations require an endpoint URL. For authenticated providers, reference a saved server_credential_id inside the component configuration or supply organization credential_ids. Inline credentials are intended for short-lived or server-managed setups; do not expose them in browser code. Use fallback_providers to provide ordered runtime fallbacks. Each entry has the configuration appropriate to its component (provider plus model, voice_id, and/or language). The first configured provider is primary and fallbacks are tried in order.

​
Voice and listening behavior

​
Voice style

voice.character_profile can be neutral, warm_host, calm_professional, playful_guide, empathetic_support, or confident_expert. It controls the assistant’s delivery guidance, not a provider voice ID. For supported expressive voices, set voice.expressive_delivery to control speech_steering.pace (slow, normal, or fast), disfluencies, permitted nonverbal sounds, style (restrained, warm, or playful), and optional tts_instructions_append (maximum 4,000 characters). Sensitive, financial, safety, health, legal, error, and complaint conversations are always constrained to restrained delivery.

​
Backchannel acknowledgements

Set voice.backchanneling to let the assistant play brief acknowledgement clips while the caller is speaking:
frequency is minimal, natural, or expressive. Cues are localized for supported voice languages where a matching pack exists. They are audio-only acknowledgements and are not added to the transcript or model context.

​
Speaking and interruption plans

start_speaking_plan controls when the assistant replies after a caller stops speaking. Use turn_detection_mode (audio or vad), optional turn_detector_version (v1 or v1-mini), wait_seconds, and smart_endpointing_mode (off or livekit). livekit uses dynamic endpointing; with off, wait_seconds controls the fixed minimum delay. stop_speaking_plan controls interruption sensitivity with num_words, voice_seconds, and optional backoff_seconds.

​
Transcription vocabulary

Pass transcriber.vocabulary as an array of names, product terms, or phrases. The runtime removes blank and duplicate values before forwarding them.

​
Audio environment

Set background_sound to office, off, or a custom HTTPS audio URL. Use background_sound_volume from 0.0 to 1.0 (default 0.8). This is separate from the selected speaking voice. Controls shown as unavailable in the dashboard are not active configuration options.

​
Post-call output

  • summary_enabled enables a post-call summary.
  • analysis_enabled enables the post-call analysis pipeline; use analysis_profile_id to select an organization profile.
  • analysis_plan can request a summary, structured data, and/or success evaluation. See Create Assistant for the payload shape.
Analysis and summary output depend on retained call text. If compliance disables transcript retention, downstream artifacts and webhook fields may be omitted or redacted.

​
Webhook callbacks

An assistant server config sets a callback destination for selected server_messages. This is distinct from organization outbound webhook subscriptions. See Webhook event contracts for supported event names and payloads.

​
Security and compliance

Use the Security & Compliance section to configure AI disclosure, recording consent, transcript retention, and PII redaction. Compliance controls are active only when compliance_plan.enabled is true. See Create Assistant for the complete runtime contract and limits.