1. Inspect the live key
Create apyai_live_ key in the console with only the scopes your application
needs, then inspect the exact policy applied to it:
environmentislive.scopescontains every endpoint you will call.limits.rps,limits.burst, andlimits.concurrencycover expected load.- The credit and per-key budget posture is ready for traffic.
402 credit_exhausted until funding or
graduated signup credit is available. Sandbox keys skip this gate and therefore
do not prove live billing readiness. Do not retry a 402; fix the underlying
credit, budget, or quota condition.
2. Confirm region and latency
PyAI currently serves from one US region. There is no India or APAC data-residency option today. Put your media server, framework worker, or backend near the PyAI region when latency matters, and measure the complete path from your deployment rather than reusing an in-region benchmark. If a single US processing region does not satisfy your residency or continuity requirements, stop before production and confirm an enterprise plan or roadmap fit. See Reliability, latency and regions.3. Confirm language behavior
Language support differs by product:- Hear sync uses
language; async jobs usetranscription_language. Omit the field for automatic detection. Streaming accepts omission orlanguage=auto. Send a supported code for known monolingual audio; validate mixed-language recordings separately. Async results do not return the detected language code. Useoutput_script: "latin"for Romanized async text. - Omni availability is staged by language. Inspect
configured.language_activeandconfigured.language_fallback. - Speak language availability depends on the selected voice and delivery mode.
4. Design reconnect and idempotency
Realtime sockets do not resume:- A dropped Omni or Hear socket ends that session.
- Reconnect creates a new session.
- Re-send Omni
configureafter reconnect. - Keep your correlation id and essential application state in your backend.
session_label when you need separate call records, or deliberately
reuse your own opaque correlation value when your application can distinguish
attempts.
For REST jobs and durable writes, send the documented Idempotency-Key and
reuse it only for the identical logical operation. Honor Retry-After on
retryable 429 responses. Never parse human-readable error text; branch on the
stable error code or terminal job status.
Read the STT and TTS recovery contract
before relying on a backup. For multi-utterance Speak calls, keep one
speech_session_id and voice for the entire call.
5. Choose the data-retention path
Retention depends on the surface. In particular:- Streaming Hear audio is relayed and not written to durable storage.
- Async Hear URL input is normally transient; enabled recovery can retain private input chunks. Completed recovery inputs are deleted, and unresolved inputs expire within 7 days. The job and result are retained.
- Async uploaded audio is retained for up to 7 days; results for up to 30 days.
- Omni recordings are off by default. When enabled, they are retained for up to 30 days.
- Omni call records and Recap records are retained for up to 90 days.
store: false option and its DELETE operation is
cancellation, not erasure. Review Security and data before
handling sensitive data.
6. Exercise failure paths
Before launch, test:- Invalid key: expect
401 unauthorizedand stop. - Missing scope: expect
403 forbiddenand stop. - Empty live balance or key budget: expect a branchable
402. - Rate or concurrency exhaustion: expect
429plusRetry-After. - Realtime disconnect: reconnect as a new session and restore application context.
- Webhook replay: verify the signature timestamp and deduplicate the event id.
- Downstream timeout: prove your application fails safely when a tool, knowledge callback, or CRM destination is unavailable.
7. Monitor the production surface
Subscribe to status.pyai.com and record the request, job, call, or session identifier needed for support. Log status codes, stable error codes, retry counts, and latency. Do not log API keys, webhook secrets, raw authorization headers, or sensitive transcript content by default. Operational availability targets are not contractual SLAs. If you require service credits, a data-processing agreement, named subprocessors, or zero-retention terms, complete the corresponding enterprise agreement before launch. The no-training warranty applies by default to customer content and generated outputs; no enterprise agreement or API opt-out flag is required.Errors and limits
Stable codes, retries, idempotency, and key-specific limits.
Security and data
Retention, tenancy, and data-use boundaries.
Language support
Serving status and fallback behavior by product.
Reliability and regions
Region, latency, reconnect, status, and SLA posture.