Metrics · Pinecall Docs

Every field livekit-agents 1.8 measures, block by block, with its unit and who measures it. Times are seconds unless the unit says otherwise; timestamps are unix seconds; a field the measuring side could not fill is absent, never zero. The story of how they join is in README.md; the models are pinecall/wire/metrics.py.

block	field	unit	required	measured by	meaning
`Metadata`	`model_name`	name	no	the provider plugin	The model's name as the provider plugin reports it.
`Metadata`	`model_provider`	name	no	the provider plugin	The provider's name as the plugin reports it: anthropic, elevenlabs, soniox.
`LLMMetrics`	`type`	tag	yes	the session's llm node	livekit's tag for this block.
`LLMMetrics`	`label`	name	yes	the session's llm node	The node that measured: the plugin's label for its LLM.
`LLMMetrics`	`request_id`	id	yes	the session's llm node	The provider's id for this request, for correlating with their logs.
`LLMMetrics`	`timestamp`	unix s	yes	the session's llm node	When the request completed, unix seconds.
`LLMMetrics`	`duration`	s	yes	the session's llm node	The whole request, first byte sent to last token received, seconds.
`LLMMetrics`	`ttft`	s	yes	the session's llm node	Time to first token, seconds. -1 when the reply produced no token.
`LLMMetrics`	`cancelled`	flag	yes	the session's llm node	True when the request was cut short, usually because the caller interrupted.
`LLMMetrics`	`completion_tokens`	tokens	yes	the session's llm node	Tokens the model wrote, reasoning included.
`LLMMetrics`	`prompt_tokens`	tokens	yes	the session's llm node	Tokens the model read, cache hits included.
`LLMMetrics`	`prompt_cached_tokens`	tokens	yes	the session's llm node	Of the prompt tokens, how many were served from the provider's cache.
`LLMMetrics`	`cache_creation_tokens`	tokens	no	the session's llm node	Prompt tokens written into the cache on this request (Anthropic cache writes). 0 when the provider does not report it.
`LLMMetrics`	`reasoning_tokens`	tokens	no	the session's llm node	Completion tokens spent on hidden reasoning. Already inside completion_tokens; never add it again.
`LLMMetrics`	`total_tokens`	tokens	yes	the session's llm node	prompt_tokens plus completion_tokens, as the provider counts them.
`LLMMetrics`	`tokens_per_second`	tokens/s	yes	the session's llm node	completion_tokens over the streaming window.
`LLMMetrics`	`speech_id`	id	no	the session's llm node	The reply this request served, joining it to turn.agent. Null for a request outside a reply.
`LLMMetrics`	`metadata`	object	no	the session's llm node	The model and provider.
`STTMetrics`	`type`	tag	yes	the session's stt node	livekit's tag for this block.
`STTMetrics`	`label`	name	yes	the session's stt node	The node that measured: the plugin's label for its STT.
`STTMetrics`	`request_id`	id	yes	the session's stt node	The provider's id for this request.
`STTMetrics`	`timestamp`	unix s	yes	the session's stt node	When the request completed, unix seconds.
`STTMetrics`	`duration`	s	yes	the session's stt node	The request's length in seconds. 0 when the STT streams, since a stream has no single request.
`STTMetrics`	`audio_duration`	s	yes	the session's stt node	How much audio was pushed to the recognizer, seconds.
`STTMetrics`	`input_tokens`	tokens	no	the session's stt node	Audio tokens in, for providers that bill by token. 0 otherwise.
`STTMetrics`	`output_tokens`	tokens	no	the session's stt node	Text tokens out, for providers that bill by token. 0 otherwise.
`STTMetrics`	`streamed`	flag	yes	the session's stt node	True when the recognizer ran over a websocket stream rather than one request per utterance.
`STTMetrics`	`acquire_time`	s	no	the session's stt node	Seconds spent getting a connection before the audio could flow. Websocket providers only.
`STTMetrics`	`connection_reused`	flag	no	the session's stt node	True when the connection came from the pool rather than a fresh handshake.
`STTMetrics`	`metadata`	object	no	the session's stt node	The model and provider.
`TTSMetrics`	`type`	tag	yes	the session's tts node	livekit's tag for this block.
`TTSMetrics`	`label`	name	yes	the session's tts node	The node that measured: the plugin's label for its TTS.
`TTSMetrics`	`request_id`	id	yes	the session's tts node	The provider's id for this request.
`TTSMetrics`	`timestamp`	unix s	yes	the session's tts node	When the request completed, unix seconds.
`TTSMetrics`	`ttfb`	s	yes	the session's tts node	Time to first byte of audio after the text was sent, seconds.
`TTSMetrics`	`duration`	s	yes	the session's tts node	The whole request, seconds.
`TTSMetrics`	`audio_duration`	s	yes	the session's tts node	How much audio came back, seconds of playback.
`TTSMetrics`	`cancelled`	flag	yes	the session's tts node	True when synthesis was cut short, usually by an interruption.
`TTSMetrics`	`characters_count`	chars	yes	the session's tts node	Characters synthesized, what character-billed providers charge for.
`TTSMetrics`	`input_tokens`	tokens	no	the session's tts node	Text tokens in, for token-billed TTS. 0 otherwise.
`TTSMetrics`	`output_tokens`	tokens	no	the session's tts node	Audio tokens out, for token-billed TTS. 0 otherwise.
`TTSMetrics`	`streamed`	flag	yes	the session's tts node	True when the audio streamed back over a websocket.
`TTSMetrics`	`acquire_time`	s	no	the session's tts node	Seconds spent getting a connection first. Websocket providers only.
`TTSMetrics`	`connection_reused`	flag	no	the session's tts node	True when the connection came from the pool.
`TTSMetrics`	`segment_id`	id	no	the session's tts node	Which sentence segment of the reply this was.
`TTSMetrics`	`speech_id`	id	no	the session's tts node	The reply this audio belongs to, joining it to turn.agent.
`TTSMetrics`	`metadata`	object	no	the session's tts node	The model and provider.
`VADMetrics`	`type`	tag	yes	the VAD plugin	livekit's tag for this block.
`VADMetrics`	`label`	name	yes	the VAD plugin	The VAD plugin's label.
`VADMetrics`	`timestamp`	unix s	yes	the VAD plugin	When the report was cut, unix seconds.
`VADMetrics`	`idle_time`	s	yes	the VAD plugin	Seconds the detector spent waiting for audio in this window.
`VADMetrics`	`inference_duration_total`	s	yes	the VAD plugin	Seconds spent inside the model in this window.
`VADMetrics`	`inference_count`	count	yes	the VAD plugin	How many frames the model scored in this window.
`VADMetrics`	`metadata`	object	no	the VAD plugin	The model and provider.
`EOUMetrics`	`type`	tag	yes	the session's turn detection	livekit's tag for this block.
`EOUMetrics`	`timestamp`	unix s	yes	the session's turn detection	When the turn was closed, unix seconds.
`EOUMetrics`	`end_of_utterance_delay`	s	yes	the session's turn detection	Seconds from the VAD hearing silence to the decision that the turn is over. 0 when no end of speech was detected.
`EOUMetrics`	`transcription_delay`	s	yes	the session's turn detection	Seconds from the end of speech to the final transcript. 0 when no end of speech was detected.
`EOUMetrics`	`on_user_turn_completed_delay`	s	yes	the session's turn detection	Seconds the app's on_user_turn_completed hook took.
`EOUMetrics`	`speech_id`	id	no	the session's turn detection	The reply this turn triggered, joining it to turn.user and turn.agent.
`EOUMetrics`	`metadata`	object	no	the session's turn detection	The model and provider.
`EOTInferenceMetrics`	`type`	tag	yes	the end-of-turn model	livekit's tag for this block.
`EOTInferenceMetrics`	`timestamp`	unix s	yes	the end-of-turn model	When the prediction came back, unix seconds.
`EOTInferenceMetrics`	`total_duration`	s	yes	the end-of-turn model	Seconds from the earliest audio in the inference to the answer.
`EOTInferenceMetrics`	`detection_delay`	s	yes	the end-of-turn model	Seconds from the latest audio in the inference to the answer.
`EOTInferenceMetrics`	`prediction_duration`	s	yes	the end-of-turn model	Seconds the model itself took.
`EOTInferenceMetrics`	`num_requests`	count	no	the end-of-turn model	How many requests one inference needed. Usually 1.
`EOTInferenceMetrics`	`metadata`	object	no	the end-of-turn model	The model and provider.
`InterruptionMetrics`	`type`	tag	yes	the interruption detector	livekit's tag for this block.
`InterruptionMetrics`	`timestamp`	unix s	yes	the interruption detector	When the inference came back, unix seconds.
`InterruptionMetrics`	`total_duration`	s	yes	the interruption detector	Round trip of the latest inference, seconds.
`InterruptionMetrics`	`prediction_duration`	s	yes	the interruption detector	The model's own time on the latest inference, seconds.
`InterruptionMetrics`	`detection_delay`	s	yes	the interruption detector	Seconds from the onset of speech to the final prediction, latest inference.
`InterruptionMetrics`	`num_interruptions`	count	yes	the interruption detector	Interruptions detected so far in the session.
`InterruptionMetrics`	`num_backchannels`	count	yes	the interruption detector	Backchannels ('mm-hm', 'sí') detected so far and not treated as interruptions.
`InterruptionMetrics`	`num_requests`	count	yes	the interruption detector	Requests sent to the detector so far.
`InterruptionMetrics`	`metadata`	object	no	the interruption detector	The model and provider.
`RealtimeCachedTokenDetails`	`audio_tokens`	tokens	no	the realtime plugin	Cached audio tokens.
`RealtimeCachedTokenDetails`	`text_tokens`	tokens	no	the realtime plugin	Cached text tokens.
`RealtimeCachedTokenDetails`	`image_tokens`	tokens	no	the realtime plugin	Cached image tokens.
`RealtimeInputTokenDetails`	`audio_tokens`	tokens	no	the realtime plugin	Audio tokens read.
`RealtimeInputTokenDetails`	`text_tokens`	tokens	no	the realtime plugin	Text tokens read.
`RealtimeInputTokenDetails`	`image_tokens`	tokens	no	the realtime plugin	Image tokens read.
`RealtimeInputTokenDetails`	`cached_tokens`	tokens	no	the realtime plugin	Of all input tokens, how many came from the cache.
`RealtimeInputTokenDetails`	`cached_tokens_details`	object	no	the realtime plugin	The cached tokens by kind, when the provider breaks them out.
`RealtimeOutputTokenDetails`	`text_tokens`	tokens	no	the realtime plugin	Text tokens produced.
`RealtimeOutputTokenDetails`	`audio_tokens`	tokens	no	the realtime plugin	Audio tokens produced.
`RealtimeOutputTokenDetails`	`image_tokens`	tokens	no	the realtime plugin	Image tokens produced. Realtime models no longer emit these; kept because the library keeps it.
`RealtimeModelMetrics`	`type`	tag	yes	the realtime plugin	livekit's tag for this block.
`RealtimeModelMetrics`	`label`	name	no	the realtime plugin	The realtime plugin's label.
`RealtimeModelMetrics`	`request_id`	id	yes	the realtime plugin	The provider's id for the response.
`RealtimeModelMetrics`	`timestamp`	unix s	yes	the realtime plugin	When the response was created, unix seconds.
`RealtimeModelMetrics`	`duration`	s	no	the realtime plugin	Seconds from response created to response done.
`RealtimeModelMetrics`	`session_duration`	s	no	the realtime plugin	Seconds the session connection has been open, for providers that bill by session time.
`RealtimeModelMetrics`	`ttft`	s	no	the realtime plugin	Seconds to the first audio token. -1 when no audio token was sent.
`RealtimeModelMetrics`	`cancelled`	flag	no	the realtime plugin	True when the response was cut short.
`RealtimeModelMetrics`	`input_tokens`	tokens	no	the realtime plugin	Tokens read, text and audio together.
`RealtimeModelMetrics`	`output_tokens`	tokens	no	the realtime plugin	Tokens produced, text and audio together.
`RealtimeModelMetrics`	`total_tokens`	tokens	no	the realtime plugin	input_tokens plus output_tokens.
`RealtimeModelMetrics`	`tokens_per_second`	tokens/s	no	the realtime plugin	Output tokens over the response's duration.
`RealtimeModelMetrics`	`input_token_details`	object	yes	the realtime plugin	What a realtime model read, by kind, with the cached part broken out.
`RealtimeModelMetrics`	`output_token_details`	object	yes	the realtime plugin	What a realtime model produced, by kind.
`RealtimeModelMetrics`	`acquire_time`	s	no	the realtime plugin	Seconds spent getting a connection first.
`RealtimeModelMetrics`	`connection_reused`	flag	no	the realtime plugin	True when the connection came from the pool.
`RealtimeModelMetrics`	`metadata`	object	no	the realtime plugin	The model and provider.
`AvatarMetrics`	`type`	tag	yes	the avatar worker	livekit's tag for this block.
`AvatarMetrics`	`timestamp`	unix s	yes	the avatar worker	When the report was cut, unix seconds.
`AvatarMetrics`	`playback_latency`	s	no	the avatar worker	Seconds between forwarding the first audio frame to the avatar and the avatar reporting playback.
`AvatarMetrics`	`session_started_time`	unix s	no	the avatar worker	When the avatar session started, unix seconds.
`AvatarMetrics`	`avatar_joined_time`	unix s	no	the avatar worker	When the avatar participant joined the room with its video track, unix seconds.
`AvatarMetrics`	`metadata`	object	no	the avatar worker	The model and provider.
`TurnMetadata`	`model_name`	name	no	the session, on the ChatMessage	The model's name as the plugin reports it.
`TurnMetadata`	`model_provider`	name	no	the session, on the ChatMessage	The provider's name as the plugin reports it.
`UserTurnMetrics`	`started_speaking_at`	unix s	no	the session, on the user ChatMessage	When the caller began speaking, unix seconds.
`UserTurnMetrics`	`stopped_speaking_at`	unix s	no	the session, on the user ChatMessage	When the caller stopped speaking, unix seconds.
`UserTurnMetrics`	`transcription_delay`	s	no	the session, on the user ChatMessage	Seconds from the end of speech to the final transcript.
`UserTurnMetrics`	`end_of_turn_delay`	s	no	the session, on the user ChatMessage	Seconds from the end of speech to the decision that the turn was over.
`UserTurnMetrics`	`on_user_turn_completed_delay`	s	no	the session, on the user ChatMessage	Seconds the app's on_user_turn_completed hook took.
`UserTurnMetrics`	`stt_metadata`	object	no	the session, on the user ChatMessage	Which model handled one leg of a turn.
`AgentTurnMetrics`	`started_speaking_at`	unix s	no	the session, on the assistant ChatMessage	When the first audio frame of the reply went out, unix seconds.
`AgentTurnMetrics`	`stopped_speaking_at`	unix s	no	the session, on the assistant ChatMessage	When the last audio frame of the reply went out, unix seconds.
`AgentTurnMetrics`	`llm_node_ttft`	s	no	the session, on the assistant ChatMessage	Seconds until the llm node returned its first token.
`AgentTurnMetrics`	`llm_node_tps`	tokens/s	no	the session, on the assistant ChatMessage	Output tokens per second over the streaming window. Absent for a reply that arrived in one chunk.
`AgentTurnMetrics`	`llm_node_ttfs`	s	no	the session, on the assistant ChatMessage	Seconds from generation start until the first sentence reached the TTS. Absent when no audio came from a LiveKit TTS.
`AgentTurnMetrics`	`tts_node_ttfb`	s	no	the session, on the assistant ChatMessage	Seconds until the tts node returned its first audio chunk, counted from the first text token.
`AgentTurnMetrics`	`playback_latency`	s	no	the session, on the assistant ChatMessage	Seconds between forwarding the first audio frame and the output reporting playback. Near zero on a room; meaningful with an avatar.
`AgentTurnMetrics`	`e2e_latency`	s	no	the session, on the assistant ChatMessage	Seconds from the caller finishing to the agent starting to answer. The number a caller feels.
`AgentTurnMetrics`	`provider_request_ids`	ids	no	the session, on the assistant ChatMessage	The provider-side request ids behind this reply, for correlating with their logs.
`AgentTurnMetrics`	`llm_metadata`	object	no	the session, on the assistant ChatMessage	Which model handled one leg of a turn.
`AgentTurnMetrics`	`tts_metadata`	object	no	the session, on the assistant ChatMessage	Which model handled one leg of a turn.
`LLMModelUsage`	`type`	tag	yes	the session's usage collector	livekit's tag for this row.
`LLMModelUsage`	`provider`	name	yes	the session's usage collector	The provider: anthropic, openai.
`LLMModelUsage`	`model`	name	yes	the session's usage collector	The model name.
`LLMModelUsage`	`input_tokens`	tokens	no	the session's usage collector	All tokens read.
`LLMModelUsage`	`input_cached_tokens`	tokens	no	the session's usage collector	Of the input, tokens served from cache.
`LLMModelUsage`	`input_cache_creation_tokens`	tokens	no	the session's usage collector	Of the input, tokens written into the cache.
`LLMModelUsage`	`input_audio_tokens`	tokens	no	the session's usage collector	Audio tokens read, multimodal models only.
`LLMModelUsage`	`input_cached_audio_tokens`	tokens	no	the session's usage collector	Cached audio tokens read.
`LLMModelUsage`	`input_text_tokens`	tokens	no	the session's usage collector	Text tokens read, when the provider breaks input out by kind.
`LLMModelUsage`	`input_cached_text_tokens`	tokens	no	the session's usage collector	Cached text tokens read.
`LLMModelUsage`	`input_image_tokens`	tokens	no	the session's usage collector	Image tokens read, multimodal models only.
`LLMModelUsage`	`input_cached_image_tokens`	tokens	no	the session's usage collector	Cached image tokens read.
`LLMModelUsage`	`output_tokens`	tokens	no	the session's usage collector	All tokens written, reasoning included.
`LLMModelUsage`	`output_audio_tokens`	tokens	no	the session's usage collector	Audio tokens written, multimodal models only.
`LLMModelUsage`	`output_text_tokens`	tokens	no	the session's usage collector	Text tokens written, when the provider breaks output out by kind.
`LLMModelUsage`	`output_reasoning_tokens`	tokens	no	the session's usage collector	Tokens spent on hidden reasoning. Already inside output_tokens.
`LLMModelUsage`	`session_duration`	s	no	the session's usage collector	Seconds of session connection, for providers that bill by session time.
`TTSModelUsage`	`type`	tag	yes	the session's usage collector	livekit's tag for this row.
`TTSModelUsage`	`provider`	name	yes	the session's usage collector	The provider: elevenlabs.
`TTSModelUsage`	`model`	name	yes	the session's usage collector	The model name.
`TTSModelUsage`	`input_tokens`	tokens	no	the session's usage collector	Text tokens in, for token-billed TTS.
`TTSModelUsage`	`output_tokens`	tokens	no	the session's usage collector	Audio tokens out, for token-billed TTS.
`TTSModelUsage`	`characters_count`	chars	no	the session's usage collector	Characters synthesized, for character-billed TTS.
`TTSModelUsage`	`audio_duration`	s	no	the session's usage collector	Seconds of audio generated.
`STTModelUsage`	`type`	tag	yes	the session's usage collector	livekit's tag for this row.
`STTModelUsage`	`provider`	name	yes	the session's usage collector	The provider: soniox, deepgram.
`STTModelUsage`	`model`	name	yes	the session's usage collector	The model name.
`STTModelUsage`	`input_tokens`	tokens	no	the session's usage collector	Audio tokens in, for token-billed STT.
`STTModelUsage`	`output_tokens`	tokens	no	the session's usage collector	Text tokens out, for token-billed STT.
`STTModelUsage`	`audio_duration`	s	no	the session's usage collector	Seconds of audio recognized.
`InterruptionModelUsage`	`type`	tag	yes	the session's usage collector	livekit's tag for this row.
`InterruptionModelUsage`	`provider`	name	yes	the session's usage collector	The provider: livekit.
`InterruptionModelUsage`	`model`	name	yes	the session's usage collector	The model name.
`InterruptionModelUsage`	`total_requests`	count	no	the session's usage collector	Requests sent to the detector.
`EOTModelUsage`	`type`	tag	yes	the session's usage collector	livekit's tag for this row.
`EOTModelUsage`	`provider`	name	yes	the session's usage collector	The provider: livekit.
`EOTModelUsage`	`model`	name	yes	the session's usage collector	The model name.
`EOTModelUsage`	`total_requests`	count	no	the session's usage collector	Inference requests sent to the model.

Per day, by vendor and by version#

GET /v1/insights reads three of the turn's numbers across the window's days of the scope's calls (the day a call started, in UTC; one agent's alone when it names one), grouped by stage and by the vendor and model the turn's own report names: stt is UserTurnMetrics.transcription_delay under stt_metadata, with the average of the turn's transcript_confidence; llm is AgentTurnMetrics.llm_node_ttft under llm_metadata; tts is AgentTurnMetrics.tts_node_ttfb under tts_metadata. Each row has how many turns reported the stage, the median and the 95th percentile in seconds, so "the ears are 500 ms slower than yesterday" is two rows of two days, and "the ears are wrong" is the confidence falling. The vendor is model_provider as the plugin reports it, and absent when the report named none.

They are counted once per call, when it is sealed, from the call's own log (log/drift.py), never read off the log at the endpoint: a day holds the calls sealed in it, not those still going. Each stage's seconds are kept as a fixed histogram (log/_histogram.py: under 10 ms, 92 buckets each 10 % wider than the one below it up to 63 s, and one for anything slower), so the calls of a day, of an agent and of a version add bucket by bucket, and a median or p95 read off the buckets is within 5 % of the exact one. The same fold counts each judge's settled verdicts of the call, by the version of the agent's settings it ran on and the hash of the judge's question, which is what GET /v1/insights/drift compares (Console API); a call judged again (POST /v1/evals/judge/{call}?again=true) replaces its verdicts rather than adding them. pinecall-runtime drift rebuild counts them again from the log (The runtime CLI).