Small settings with a big impact — from the right confidence threshold to a clean knowledge base.
In the folder browser at /Admin/MailAccounts/Manage, folders are loaded live
from the mailbox. For a usable training corpus, both the inbox and the
sent folder must be selected — QaPairs only emerge from matching an incoming
question to the reply that was actually sent via thread association.
Common pitfall: if only the inbox is selected, the knowledge base stays empty because no matching replies can be associated.
Signature and promotional footer patterns under /Admin/CleaningPatterns should
be added before the first test import — they affect all QaPairs produced during import. A
pattern added later only applies to future imports, not retroactively.
Use the two deletion scopes: when re-testing, choose whether to delete only the derived QaPairs (raw messages remain) or everything including raw messages — handy for quickly testing cleaning patterns multiple times without re-importing each time.
The thresholds in data/retrieval-settings.json (default: high ≥ 0.85, medium ≥
0.60) determine how often replies go out automatically versus into the review queue.
Especially at the start, it pays to set the full-automation threshold higher rather than
lower, and only lower it after a few weeks of observed match quality.
Under /Admin/Curation, SupportFlowAI shows duplicate candidates above 97%
similarity, outdated entries, and recurring inquiries with weak confidence. The latter is
especially worth checking regularly — it points directly to topics the archive still lacks
good answers for.
Only the SystemAdministrator role can activate full automation, and even then only with a two-step confirmation (checkbox + re-entering the password). For day-to-day operation, most teams are well served by the AccountManager role (accounts/curation) or Reviewer (review queue only) — keeping the number of accounts able to toggle full automation deliberately small.
Use Auditor for spot checks: the read-only, PII-masked auditor role is well suited for occasional quality spot checks without accidentally viewing customer data in plain text.