Struct RealtimeSession 

pub struct RealtimeSession { /* private fields */ }

Available on crate feature realtime only.

An active realtime session.

Cheap to share indirectly via the channels it owns; call RealtimeSession::close to terminate the background task.

Implementations

impl RealtimeSession

pub async fn send_audio(&self, pcm: Bytes) -> ZaiResult<()>

Send raw 16-bit little-endian mono PCM.

With InputAudioFormat::Wav (the default), the bytes are wrapped in a 16 kHz WAV container. With Pcm16 or Pcm24, they are sent as raw PCM and the selected format declares the corresponding sample rate.

pub async fn send_video_frame(&self, jpeg: Bytes) -> ZaiResult<()>

Upload a JPEG frame for passive-video mode.

pub async fn commit_audio(&self) -> ZaiResult<()>

Commit buffered audio for inference in client-VAD mode. Server-VAD commits automatically and normally does not need this command.

pub async fn clear_audio(&self) -> ZaiResult<()>

Clear audio buffered by the server without triggering inference.

pub async fn send_text(&self, text: impl Into<String>) -> ZaiResult<()>

Inject a user text message into the conversation history.

pub async fn send_function_output( &self, call_name: impl Into<String>, output: impl Into<String>, ) -> ZaiResult<()>

Feed back a function-call result.

pub async fn delete_item(&self, item_id: impl Into<String>) -> ZaiResult<()>

Delete an item from the server-side conversation history.

pub async fn retrieve_item(&self, item_id: impl Into<String>) -> ZaiResult<()>

Ask the server to emit the current representation of one conversation item via ServerEvent::ConversationItemRetrieved.

pub async fn create_response(&self) -> ZaiResult<()>

Trigger model inference (response.create).

pub async fn cancel(&self) -> ZaiResult<()>

Cancel the in-flight response (response.cancel), e.g. on interruption.

pub fn events( &self, ) -> Pin<Box<dyn Stream<Item = ZaiResult<ServerEvent>> + Send + '_>>

Stream of server metadata events (transcripts, response lifecycle, errors, heartbeats). Audio deltas are decoded only onto Self::audio_stream to avoid retaining a second base64 copy. The first subscriber receives events buffered since session creation; later subscribers start at the live tail. A lagged consumer or background session failure is surfaced as an error instead of silently losing protocol events.

pub fn audio_stream( &self, ) -> Pin<Box<dyn Stream<Item = ZaiResult<RealtimeAudioChunk>> + Send + '_>>

Stream of decoded 24 kHz, mono, 16-bit PCM output chunks.

Lag is an error because dropping a PCM chunk would silently corrupt the resulting audio stream.

pub fn model_name(&self) -> &str

The model id sent in the initial session.update event.

pub async fn request_close(&self) -> ZaiResult<()>

Signal the background task to close without awaiting it.

Use when the session is shared (e.g. behind an Arc), driven from a tokio::select!, or closed reactively on a shutdown signal. This only signals a dedicated shutdown channel; the background loop observes it independently of the bounded outbound queue and exits. For deterministic, awaited teardown use RealtimeSession::close.

pub async fn close(self) -> ZaiResult<()>

Close the session and wait for the event loop to finish.