The pitfalls that actually show up in real first-time setups โ and how to avoid them from the start.
Common assumption: "The Sent folder already has the question and answer together (the reply above the quoted request), so importing just Sent should be enough." In reality, SupportFlowAI forms a question/answer pair from the first inbound message (the customer's request, from INBOX) plus the first subsequent outbound message (the reply, from Sent) in the same thread. Without INBOX, the question side of the pair is missing.
The matching is done via the mail threading headers (References,
In-Reply-To), not via the structure of the message text โ so it doesn't matter
whether the reply in the Sent folder appears above or below the quoted request. The quoted
request is automatically stripped from the reply text during cleanup anyway.
Rule of thumb: When selecting folders for training accounts, always select both INBOX and Sent (the folder name is provider-dependent, e.g. "Sent Items").
If the automatic connection test fails, the specific error message appears directly below the status badge in the account overview โ not just as a hover tooltip. Typical causes:
CPU mode is entirely sufficient for a test run and smaller mailboxes โ the models are small enough to respond in a reasonable time without a GPU. GPU mode pays off once response quality and speed at higher volume matter more than resource efficiency.
SupportFlowAI is deliberately not designed for real-time responses โ replies may well take a few minutes. Resource efficiency takes priority over latency here.
A purely RAM-based index doesn't scale indefinitely with mailbox size. As a rule of thumb:
Performance on a server with plenty of free memory, Balanced as a
good middle ground, Lean on weaker test hardware or when the knowledge base is
expected to grow significantly.
Wait up to one full background sync interval (default 5 minutes, viewable/changeable under
/Admin/MailSync), then check /Admin/SystemHealth โ there the worker
heartbeat shows whether the service is ticking at all, and whether it was intentionally
paused.
Check whether Ollama is actually running (http://localhost:11434) and that the
required models were actually pulled via ollama pull (see Installation).
Stop the application, delete data/supportflowai.db (and, if applicable,
data/dataprotection-keys/), then restart โ the first-time setup will appear
again from the beginning.