Skip to main content
What to expect operationally, and where to watch live health.

Status & incidents

  • Status page: status.pyai.com, live surface health plus posted incidents and maintenance.
  • Subscribe: machine feeds at /incidents.json, /history.json, and Atom /feed.xml; or get email updates (double opt-in) via POST https://api.pyai.com/v1/status/subscribe.
  • During an incident, the status page is the source of truth for scope, impact, and updates.

Uptime targets & SLA

PyAI runs to these operational availability targets:
These are operational targets, not a contractual commitment by themselves. A contractual SLA (with credits) is available on enterprise agreements, talk to us. Today PyAI runs a single region; higher availability tiers ride the multi-region program on the roadmap.

Latency

  • Omni (English), in-region: about 390 ms median turn-taking, measured from utterance-end to the first agent-audio frame.
  • Hear + your stack: PyAI owns the selected Hear speech-to-text leg. Your reasoning, speech output, and playback add to the end-to-end number.
Serving region. PyAI is served from a single US region today, there is no India/APAC edge yet. Callers far from the US will see cross-region round-trip added to every turn. If low latency in India is critical to your launch, raise it with us early; an India/APAC edge is roadmap, not available today.

Regions & data residency

The platform runs in one US region. There is no India (DPDP) data-residency option today. Mitigations that exist now: streaming Hear and Omni live audio are relayed, not persisted; recordings are off by default; and stored artifacts have bounded retention. For formal residency requirements, contact us, it’s an enterprise/roadmap conversation.

STT failure recovery

Recovery is enabled for all organizations on the supported paths below. PyAI remains the primary engine; a backup is attempted after a recoverable service failure or incomplete word timings. Your application continues using the same PyAI endpoint, key, and response format. Requested STT redaction applies after either engine returns. Protection failure produces an error instead of unprotected text. Protected previews are optional, so clients must also handle finals-only streams. A temporary verifier outage can leave an explicit gap while subsequent protected speech continues. See live redaction. Timeouts, connection loss, retryable service errors, and unusable responses can trigger recovery. Async jobs and verbose_json uploads also check that word timings cover the transcript and stay within the recording. Recovery never invents timestamps; if it cannot return complete timings, the request fails. Live finals can be text-only. When a live final includes word times, they are seconds from the beginning of that stream. If only word timing is incomplete, the default strict policy ends with word_timings_failed. The optional best_effort policy reports partial timing coverage in the final usage frame and closes normally. See live word timestamps. Incorrect but valid words do not trigger a second model or an accuracy judge. Authentication and invalid-request errors must be corrected by the caller. Live recovery can replace a displayed partial. Append corrected final events once per utterance_id; do not append both speech_final and final as separate sentences. If both engines fail, the replay limit is exceeded, or the audio boundary cannot be recovered safely, Hear reports recognition_interrupted and closes the stream. Keep the finalized transcript and mark the session incomplete. A dropped customer WebSocket still requires a new session. Async recovery uses the original job deadline across attempts. The recovery budget is 160 seconds from job creation; this is a processing limit, not a completion-time SLA or a throughput guarantee. A failed job emits the normal failure callback. See recordings and callbacks and data retention.

TTS: keep one voice throughout a call

TTS recovery is selected before the call’s first audio response. Once selected, the engine and voice stay fixed for that call. A backup voice can sound different from the requested stock voice, but it cannot replace that voice halfway through the conversation. If the selected engine later fails, the request fails; a retry keeps the same selection. Never splice another voice into a partially played reply. For enabled Speak organizations, pass one unique speech_session_id on every utterance and retry. Keep voice unchanged:
Use the same ID for the next reply in this call and a new ID for the next call. It is scoped to your organization and project, lasts at most 24 hours, and is separate from request idempotency. Do not reuse it after a call ends. Omitting it uses ordinary Speak synthesis without call-level external recovery. Recovery supports qualified stock voices with neutral delivery. Custom/cloned voices and directed emotions are excluded. For a new integration, select en1 (English), es1 (Spanish), fr1 (French), de1 (German), or hi1 (Hindi). Do not assume every voice in the catalog supports recovery. An unavailable or unqualified recovery configuration returns an explicit error. A storage failure also fails explicitly because PyAI cannot safely confirm the call’s voice selection. Responses identify the selected profile with x-pyai-speech-profile: primary or recovery. Backup Speak audio is returned as a complete buffered render and reports x-pyai-stream: buffered; allow for higher time to first audio. Native streams retain their normal delivery mode. A 409 voice_session_selection_conflict means concurrent first requests selected different profiles: retry with the same ID. A mismatched voice or expired session also returns 409. Enabled Omni voice-agent calls track this selection within their existing call session, including greeting, cached audio, fillers, and later replies. Applications do not add speech_session_id to Omni configuration. A new Omni connection is a new call; it is not a way to resume a previous voice selection.

Reconnect & retry

There is no mid-call session resume: a dropped realtime socket ends the session, and reconnecting starts a new one (for Omni, re-send configure). Recommended handling:
  • Hear can reject the upgrade with HTTP 429 or close a busy socket with 1013. Honor Retry-After when available and use exponential backoff with jitter. For 4429 (concurrency cap), release a slot before retrying.
  • After 1011, retain accepted results and mark the session incomplete. For 4401 or 4403, correct the key or scope before reconnecting.
  • Keep per-call state (your session_label, persona/context, a short running summary) in your backend so a reconnect can re-prime the session and continue gracefully.
See the Omni protocol reference for the full close-code table and reconnect semantics.

See also

Errors & limits

Error codes, rate, and concurrency limits.

Security & data

Tenancy, retention windows, and recordings.

Omni wire protocol

Close codes and reconnect semantics.

Pricing

Current rates, included usage, and plan availability.