Skip to main content
Keep visual analysis, OCR/extraction and chat on explicit queues. Run only the workers for the features you enabled. All examples assume a configured Redis connection; use your application’s actual asynchronous connection name. For example, configure separate workers with finite attempts and timeouts:
Set the selected connection’s retry-after greater than the largest worker timeout and applicable lock lease; for these examples it must exceed 600 seconds. If other jobs require a shorter retry interval, give this work its own Laravel queue connection rather than changing a shared connection blindly. Configure finite storage/provider network timeouts as well. A sync connection is not a background worker. Automatic analysis/extraction intentionally does not run paid processing inside an upload request; the assistant is unavailable without a supported asynchronous connection.

Admission and cost control

  • max_pending_per_workspace bounds outstanding visual/content admission; the shipped default is 100 for each subsystem.
  • Source byte/pixel/duration/frame/page caps bound individual inputs.
  • Assistant max_active_runs limits active responses; model usage_weight is a relative display/usage weight, not a monetary budget.
  • Provider quotas and spending limits belong to your provider account and deployment.
A timed-out HTTP request may already have been billed. Ready identical version/profile results are reused where implemented, but an explicit retry after an ambiguous failure can repeat charges. Automatic analysis has no durable exactly-once billing outbox or automatic whole-library backfill. Do not describe those as delivered features.

Persistence and cancellation

Queued/running/result states survive an ordinary browser reload. A worker or network failure is different: a killed process may require operator intervention or a new explicit attempt. The assistant checks for stale active runs on subsequent conversation access; it does not guarantee immediate recovery at the moment of a server crash. Stop cannot undo a committed mutation. Use assistant history, retention and scheduled maintenance. File deletion/replacement makes stale versions ineligible for publication. Engine synchronization and retained SQL projections have their own lifecycles; disabling a switch is not bulk deletion.

Diagnose by subsystem

Use configuration defaults and the relevant capability guide before increasing any limit. Keep provider credentials, raw document content and private attachments out of public logs. Encryption at rest does not make external AI processing local: selected inputs must be available to the configured provider.