Manage Profile AI

The Profile AI administration section combines system-wide generation configuration with privacy-safe queue and weekly-sweep monitoring. Only a CareMap system administrator can open it or change its settings.

Configure generation

Open Admin > Profile AI to review or change:

• the provider model and global system prompt;
• the maximum simultaneous category generations;
• weekly scheduling and queue lease settings;
• evidence, context, and output limits; and
• retry and provider timeout controls.

Saving a changed model or prompt creates a new immutable configuration version. Already queued category jobs keep the version they captured, so a mid-sweep edit does not silently change their instructions. Keep prompts general and never paste a person's private care content into global configuration.

The default deployment-wide capacity is four. Raising it increases provider traffic and credit-reservation pressure across all profiles. Treat a capacity change as an operational change and watch queue depth, oldest-job age, provider limits, and database load afterward.

Monitor a weekly sweep

The overview refreshes automatically and shows active slots, queued and retry work, terminal outcomes, throughput, high-water depth, and the oldest queued job. Sweep cards show elapsed time, people and categories processed, remaining work, categories per minute, and low-credit or failed outcomes.

One category can retry or fail without discarding successful sibling categories. Low credits stop only the affected person's remaining work. A sweep can continue after Sunday and resumes from durable database state after a server restart.

Alerts identify conditions such as an old backlog, a sweep outside its service window, repeated lease expiry or rate limits, provider authentication failure, high terminal-failure or low-credit rates, and unsettled provider work. Use the support list to locate a run by opaque person, run, job, or sweep ID and to review statuses, attempts, timing, model metadata, and settled credits.

The dashboard and operational logs deliberately omit names, filenames, prompts, evidence, generated summaries, citation titles, and provider output. Do not copy private source content into incident notes.

Respond safely

For a queue incident:

1. Confirm the current alert and whether new work is still finishing.
2. Check active capacity, oldest queued age, retry timing, sweep progress, and provider-failure signals.
3. Use the support list's opaque identifiers to distinguish one person run from a deployment-wide failure.
4. Correct provider credentials or capacity only through approved deployment controls; never edit queue rows by hand.
5. Verify that leases recover, backlog age falls, settlements finish, and the sweep reaches a terminal state.

Pausing weekly scheduling prevents new scheduled sweeps; it does not erase durable queued work. The Force sync now profile action is available only in development builds and production rejects a constructed request.

See the repository's docs/PROFILE_AI_OPERATIONS.md runbook for deployment, alert thresholds, incident steps, and recovery details maintained for operators.