Module rtvi 

RTVI-protocol observer (pure frame→message mapping, no network).

Translates pipeline Frames into outgoing RTVI client-protocol messages (user/bot transcription, speaking edges, LLM/TTS lifecycle, function-call signalling, metrics). Mirrors pipecat’s processors/frameworks/rtvi/observer.py (RTVIObserver) and the v1 message shapes in .../rtvi/models.py.

This module is pure: it maps a frame to an RtviMessage and hands it to an injectable RtviSink. The actual transport (WebSocket/Daily/etc.) lives in the embedder; tests use the in-memory VecSink. No reqwest/network dep is pulled into flowcat-core (the networked OTel/Sentry/Langfuse exporters live in flowcat-services instead).

What is mapped (parity with pipecat RTVIObserver.on_push_frame)

Frame	RTVI message type
Frame::UserStartedSpeaking	user-started-speaking
Frame::UserStoppedSpeaking	user-stopped-speaking
Frame::BotStartedSpeaking	bot-started-speaking
Frame::BotStoppedSpeaking	bot-stopped-speaking
Frame::SttMute	user-mute-started / user-mute-stopped
Frame::Transcription (user_id != "bot")	user-transcription (final: true)
Frame::Transcription (user_id == "bot", realtime s2s)	bot-transcription
Frame::InterimTranscription (user_id != "bot")	user-transcription (final: false)
Frame::LlmResponseStart	bot-llm-started
Frame::LlmResponseEnd	bot-llm-stopped
Frame::LlmText	bot-llm-text (+ bot-transcription on sentence end)
Frame::TtsStarted	bot-tts-started
Frame::TtsStopped	bot-tts-stopped
Frame::TtsText	bot-tts-text + bot-output (spoken: true)
Frame::FunctionCallsStarted	llm-function-call-started (per call)
Frame::FunctionCallInProgress	llm-function-call-in-progress
Frame::FunctionCallResult	llm-function-call-stopped (cancelled: false)
Frame::FunctionCallCancel	llm-function-call-stopped (cancelled: true)
Frame::Metrics	metrics (bucketed ttfb/processing/tokens/characters)

Structs

RtviMessage

One outgoing RTVI server→client message. Serializes to the exact pipecat v1 wire shape: { "label": "rtvi-ai", "type": "<type>", "data"?: { … } }.

RtviObserver

Pure frame→RTVI-message observer (PROCESSOR-DESIGN §5.1; pipecat RTVIObserver). Plugs into the pipeline Observer fan-out via FrameObserver; for each pushed frame it emits zero or more RtviMessages to the injected RtviSink.

RtviObserverParams

Toggles for which RTVI message families the observer emits. Mirrors pipecat RTVIObserverParams. All default to on (the conventional client UX), except audio levels (off — those need the audio payload + a rate limiter).

VecSink

An in-memory RtviSink that records every message — the test/inspection sink. Cheap to clone-share via Arc.

Enums

FunctionCallReportLevel

How much function-call detail an RTVI message exposes (pipecat RTVIFunctionCallReportLevel). Defaults to None — the most secure level that still emits the event (tool-call-id only, no name/args).

Constants

MESSAGE_LABEL

The constant label on every RTVI message (pipecat MESSAGE_LABEL).

PROTOCOL_VERSION

RTVI protocol version this observer emits (mirrors pipecat models.PROTOCOL_VERSION).

Traits

RtviSink

Where the RtviObserver hands finished messages. Decouples the pure frame→message mapping from any transport — the live transport (WS/Daily) in the embedder implements this; tests use VecSink.