deflect --source docs,repo,tickets --cite always
Deflect resolves tier-one tickets from your own documentation, changelog and codebase. Every answer carries the source it came from. Anything it cannot prove, it escalates instead of guessing.
Support answers that cannot be checked get escalated by the customer anyway. Deflect shows its working, every time.
Subject: Webhook retries stopped after upgrade
We moved to v4 last Thursday and our webhook endpoint stopped receiving retries on 5xx. It still gets the first delivery. Nothing in our logs changed on our side. Did retry behaviour change?
Retry behaviour did change in v4. Retries are now opt-in per endpoint rather than on by default, and the flag does not carry over from v3.
“Webhook retries are now configured per endpoint via retry_policy. Endpoints migrated from v3 default to none.”
“Set retry_policy: exponential on the endpoint to restore v3 behaviour. Takes effect on the next delivery.”
Four stages. Any one of them can stop the pipeline and hand the ticket to a person.
Billing, account deletion and security tickets never enter the pipeline at all. They route to a human on arrival, no exceptions and no configuration that can accidentally allow it.
Searches docs, repo, changelog and previously solved tickets together, because the real answer is usually split across two of them.
Every claim in the draft is bound to a retrieved passage. A sentence that cannot be bound is removed before the reply is scored, not after.
Above threshold it replies and tags the thread. Below it, the draft plus the sources goes to an engineer as a head start rather than a blank box.
Roughly 4,100 tickets a month, developer audience, self-serve and enterprise mixed.
Support AI fails in three specific ways. Each of these exists because of one of them.
An answer that cannot be traced to an indexed source is never delivered to a customer. It becomes an escalation with the draft attached.
If the retrieved passage predates the customer's version, Deflect flags the conflict rather than confidently quoting an obsolete page.
Billing, deletion and security are excluded in the engine, not in a prompt. There is no phrasing that talks it into handling one.
It answers them blind, with citations, and we diff its replies against what your team actually sent. You will see the deflection rate and the misses in the same report.