Training ClickSort on your folders
How ClickSort learns your Outlook folder structure, why its suggestions improve, and which folders it learns from and how to reset the model.
ClickSort does not ship knowing anything about your mailbox. It learns from the mail already in your folders and from the way you file, on your device, and its suggestions get better as it goes.
How the learning works
ClickSort builds a classifier over your own folders, in two ways:
- From the mail already in your folders. The first time you open the pane, and again once the model is a day old, ClickSort reads up to 50 of the newest messages in each of your folders. From each it takes the sender, the subject and the opening text of the body (the preview Microsoft keeps, up to 255 characters).
- From every message you file. When you file, by suggestion or by search, the message's sender and subject (in batch view, its opening text as well) are added to the folder you chose. Undo takes them back out.
The next message that looks like the mail in a folder gets that folder as a suggestion. The sender counts most, then the sender's domain, then words in the subject, then words in the opening text.
What is worth knowing about how that is built:
- The model keeps words, not messages. For each folder it keeps sender addresses, sender domains and single words from subjects and body openings, each with a count. It does not keep whole messages, whole subjects or whole bodies, but the words it keeps are real words from your mail.
- It lives on your device, in the add-in's browser storage inside your Microsoft 365 session, partitioned per tenant and mailbox. It is not uploaded, and there is no server-side classifier.
- There is no LLM anywhere in the filing path. Suggestions come from a small statistical model that runs locally, which is why they are instant and why nothing has to be sent anywhere to produce them.
Getting good suggestions faster
- File the way you actually want to file. Every file is a training example, including the ones where you override the suggestion. Correcting a wrong suggestion is the most useful thing you can do.
- Give a folder a handful of examples before expecting it. A folder you have filed into once is not yet a pattern.
- Keep the folder you file to specific. ClickSort learns the distinction you make; if two folders hold the same kind of mail, no model can tell them apart.
Which folders it learns from
Every folder in your mailbox except the system folders: Inbox, Sent Items, Drafts, Deleted Items, Junk Email, Outbox, Conversation History, Scheduled and Microsoft's sync and recovery folders are skipped automatically, as are search folders. An ordinary Archive folder is included. There is not yet a setting to leave out a folder of your own.
Resetting the model
Clearing the add-in's stored data resets the model to nothing: the learned words and counts, the folder cache and your preferences all go, and ClickSort learns your folders again the next time you open the pane. Removing the add-in has the same effect. Nothing needs to be deleted from a Vonian server, because nothing was ever there.
The cached list of your folders refreshes when it is more than a day old, and is replaced rather than accumulated.
Related
- Filing an email โ the action the model learns from
- Data and privacy โ exactly what is held, where