Classification endpoints expose structured watch extractions attached to canonical messages. Orchestration may evolve between releases; always bind integrations to the published OpenAPI rather than assumed timings.

​
Typical surfaces

  • GET /v2/ai/model-profiles — Available model personas/options for routing requests.
  • PUT / DELETE /v2/ai/connectors/{connector_id}/key — Store or remove provider credentials Watchatic uses server-side.
  • GET / PUT /v2/classification-settings — Tenant tuning knobs surfaced to clients when authorized.
  • GET /v2/classifications — Listing/filter endpoint over classification rows.
  • GET /v2/classifications/{classification_id} — Single row detail.
  • POST /v2/classifications/{classification_id}/corrections — Submit human corrections that append revised structured output.
  • POST /v2/messages/{canonical_message_id}/classifications/retry — Ask the backend to re-run extraction for one message.
Terminal states, retry rules, and payload discriminators are spelled out per schema in the reference.

​
Integration stance

Poll or subscribe at the application layer after triggering retries; processing duration depends on provider latency and queue depth. For mismatches between chat content and structured fields, prefer corrections over blind retries when both apply.