LLM classifications

Every new email Saymail receives goes through a classification pass. The configured LLM reads the message and produces a small structured result that drives sort order, summaries, and a few inline UI affordances. Classification is fully transparent — you can see exactly what the LLM said for every message.

What the classifier produces

Each classification result includes:

Initial importance score — the LLM’s first estimate, on a 0–100 scale.
Final importance score — the same score after any decay (see below). For fresh email, initial and final match.
Reasoning — a short, human-readable explanation of why the LLM scored the message the way it did.
Signal tags — small labels the prompt elected to add (e.g., invoice, meeting, personal, newsletter).
Detected language — the language code the LLM identified in the original message.
Translation (optional) — a translated version of the message body. See Email translations.
Decay (optional) — a duration or date after which the importance fades. Useful for time-sensitive mail that becomes irrelevant once a deadline passes.
Suggested replies (optional) — short reply ideas the LLM proposes.

Every field except the score is optional — the prompt decides what to fill in.

Where to see classifications

Email list — emails sort by importance by default. A small bar at the start of each row shows the score visually.
Email viewer — the top of the viewer shows the score, signal tags, and the LLM’s reasoning. Below the body, the translation block and suggested replies appear when present.
Classification Details panel — the full breakdown for the selected email: every field the LLM returned, the prompt used, and the connection that handled the request. Open it from the command palette or the email’s “More” menu.

How classifications are produced

When a new email arrives (or you trigger reclassification), Saymail:

Picks the classification prompt assigned to the account, or the profile’s default if the account doesn’t have one.
Picks an LLM connection based on priority. If the prompt has a preferred connection set, that one is used; otherwise it’s the highest-priority healthy connection.
Sends the message body + prompt to the model.
Parses the JSON response into the fields above and stores the result.

The classifier runs in batches every 15 minutes by default, plus once per inbox refresh. New mail typically gets classified within seconds.

Reclassifying

You can rerun the classifier on existing email — useful after editing a prompt or switching to a better model.

One email: open the Classification Details panel and click Reclassify.
Many emails: select them in the list (Shift- or Ctrl/Cmd-click), open the bulk-action menu, and choose Reclassify. Saymail re-queues every selected message.

Reclassification doesn’t delete anything — the new score and reasoning replace the old ones.

Per-account customisation

Different accounts can use different prompts. A Work account might use a prompt that emphasises invoices and meeting requests; a Personal account might use one that demotes newsletters. Set the prompt on the account, or set a default on the profile and inherit it across all accounts in that profile.

Privacy

Classification runs through whichever LLM connection you’ve configured. A local LLM keeps email body text on your machine; cloud providers see the message contents for the duration of the request. See the privacy notes on each option for specifics.