Production checklist

Treat the provider client as infrastructure and the selected model as a tested application dependency.

Configuration

• Keep API keys and provider clients in server-only modules.
• Require either apiKey or a preconfigured SDK client at startup.
• Make completionApi explicit when a deployment depends on Responses or Chat behavior.
• Keep custom baseUrl and headers in trusted deployment configuration.
• Use an application allow-list for selectable model IDs.

Capability tests

• Smoke test the exact endpoint, adapter, and model ID.
• Test both .send() and .stream() when the UI streams output.
• Exercise required tool calls and complete streamed tool arguments.
• Validate output-schema behavior before relying on parsed product data.
• Test image, document, embedding, speech, or transcription requests with representative media.
• Treat model listing as inventory, not capability proof.

Reliability

• Bound request time, retries, and concurrency at the application boundary.
• Retry only transient failures and use idempotency around downstream writes.
• Move bulk embeddings and slow media work to durable workers.
• Keep one embedding model and vector dimension configuration per index.
• Avoid silently switching provider, adapter, or model after a failure.

Security and data

• Validate tool permissions outside model arguments.
• Remove secrets and unnecessary personal data before provider requests.
• Keep raw images and audio in object storage, not memory or traces.
• Apply access control and retention policy to transcripts and generated assets.
• Return safe application errors instead of raw provider responses.

Observability

Record the selected provider, model, adapter, latency, usage, error class, and application request ID. Attach Lens, Langfuse, or another Anvia observer when model and tool traces are needed. Choose payload capture deliberately; prompts, documents, images, transcripts, and tool results can contain sensitive data.