Skip to main content
When something common breaks, the support is covered by a wave of similar appeals, the Incidents section collects such waves into one object: it is associated with tickets, updates are sent to it, and for the time of the confirmed incident, the AI stops chasing each client on the same diagnosis. Both built-in detectors work on the same principle: automation only suggests an incident, confirms the human. The proposed incident does not change anything in the operation of the module.

Detector of similar treatments

A simple lexical detector: five or more similar calls in 30 minutes turn into an incident sentence, and you can see the signals from which the sentence is assembled: confirm not blindly.

Complaints against hosts

The second detector is smarter and catches things that synthetic checks don’t see: regional locks, provider trottling, DPI changes. The host can be green in all samples, and customers from one region are no longer connected. It works on the Observer model, which reads every move of the client (regardless of the AI mode and who is conducting the dialogue) and extracts a signal: whether it is a complaint against a particular host and from which region the client is writing, and then the module counts for each Flow the number of different complaining customers out of the window and compares with the number of active subscribers. The threshold is double: percentage plus the absolute minimum of complaining, so that in a small installation one unhappy person does not like an incident. Crossing the threshold suggests an incident by unit, writes a line to the unit’s health log (it does not change its health status), and sends a Telegram notification category “Mass Complaints.” The detector turns on by assigning the model to the role of “Observer”, there is no separate flag; the thresholds are configured in the “Mass Complaints” card. You can also turn on the clarification of the region from the client. In analytics, this detector is focused on the “Mass Host Complaints” table: for each unit, the number of complaints, the proportion and the current health of the samples. The combination of “health is normal, complaints are growing” is the main signal. Separately, the geography of complaints and calls in which the host failed to recognize: if there are many, customers call units not as they are named for you.

The life cycle of the incident

The incident, created manually, immediately receives the status of “Confirmed” Tickets are associated with the incident automatically (by cluster) or manually through “Connect Ticket” with the search. In the confirmed incident, the mailing list is available: one update goes to the customers of all the related tickets, the panel shows how many messages are actually queued. “Eliminated” related tickets does not close: some of them fell into the cluster by coincidence of words, and their problems may remain.
The active incident is also visible on the support review page: a separate card in the bottom row.