Real-time Alerts
Trigger notifications when retriever results match conditions on ingest, schedule, or demand
Why do anything?
Important content changes go unnoticed without proactive monitoring. Manual checking doesn't scale.
Why now?
Real-time applications need instant notifications when relevant content appears.
Why this feature?
Alert system with three execution modes: ON_INGEST (real-time), SCHEDULED (cron), ON_DEMAND (API trigger). Multiple notification channels.
How It Works
Alert system monitors retriever results and triggers notifications based on conditions.
Trigger
ON_INGEST, SCHEDULED, or ON_DEMAND execution
Retriever Execution
Run associated retriever query
Condition Evaluation
Check if results match alert conditions
Notification
Send alerts via EMAIL, SLACK, or WEBHOOK
Why This Approach
Multiple execution modes support different monitoring patterns. Retriever-based alerts leverage existing search logic.
Where This Is Used
Integration
alert = client.alerts.create(retriever_id="...", execution_mode="ON_INGEST", channels=[...])
Recent updates
Full changelog- Sep 27, 2026An alert's batch action runs as soon as the alert fires on uploadAn on-ingest alert can send its matches to a collection for further processing. When the alert fired at the moment an object was written, those matches were queued and waited for some unrelated batch to finish, or indefinitely in a namespace without one. They are now submitted when the alert fires.
- Sep 26, 2026Alerts fire when a match scores above 1Alert matches were validated to a score between 0 and 1, and retriever scores can land outside that range: cosine similarity on float32 returns 1.0000002 for a document matched against its own content, and BM25 has no upper bound. One such score dropped every match for the alert, so an on-ingest alert whose query was the document itself never fired. Scores are now clamped into 0 to 1 before validation.
- May 20, 2026Faster system alert evaluationAlert sweeps now read metrics in parallel and evaluate across orgs in a single pass, keeping large-tenant alert evaluation well within its window.