Currently, Stalwart supports an LLM classifier. This leans on generative text/chat completion models.
TypeSafe’s Jev takes a different approach: instead of a chat completion, you send a yes/no question (“Is this email spam?”, optionally with criteria for what counts and what doesn’t) and get back a probability. It doesn’t speak the OpenAI spec at all, so it can’t be dropped into the existing classifier config, it needs its own backend. I think this would be fantastic addition for better spam classification that would work at scale, given the speed of Jev.
There’s already someone who ran Jev against a bunch of spam e-mails from a public data set runs 19,528 emails with no training data and compares it against TF-IDF baselines trained on the dataset’s own labels. It shows a 98.33% accuracy without any prior training.
Cost is roughly $0.058 per 1,000 emails at the listed price of $0.042 per million input tokens, with output tokens free.
The caveats are in the README and worth reading: one run, one model version, one public dataset, and the criteria were tuned against that dataset’s own mistakes.
Proposal: add Jev as a second classifier backend alongside the OpenAI-compatible one, with its own config section (endpoint, API key, question/criteria definition, score mapping). The criteria should be admin-configurable rather than hardcoded — the PoC’s biggest gain came from spelling out precisely which mail counts as spam and which doesn’t, and the README explicitly notes those criteria fit that dataset and that other deployments should write their own.