The cool thing about this spam filter is that the LLM provides an explanation of why it thinks something is spam, which i can include as a local reply to them email after my workflow marks it as spam.

The workflow so far is:

1. Feed in raw email -> generate analysis of if its spam or not

2. Feed in analysis -> get LLM to generate a single SPAM or NOTSPAM token by reading the analysis and making a decision

I then check for this SPAM token and mark the email as +spam and/or +analyzed with notmuch.

the tagging potential of this beyond just spam is huge. best of all its all local AI, so no need to feed my person emails to the openai machine.

nostr:note13edqfznfkud7t48crkhu8y4h938jwftcekg7nk6hwk9366djta7s8lr7tr