2026-11-12 · America/Los_Angeles · 社区动态 · #9

Cleaning quoted replies and signatures out of email with rules before a model reads it

I build an email product for agents, and email is a bad input for a model as it arrives. A reply usually carries the thread quoted below it, plus signatures and legal footers. So each message goes through a cleaning step with no model in it, and the summarizer, the embeddings and the auto-reply all read that cleaned text. It cuts at the earliest quote boundary (an "On ... wrote:" line, an Outlook header block, the Spanish, French and German versions of those), then drops trailing ">" lines, then cuts marketing footers, then signatures. If the cuts would leave nothing, it returns the original…

热度 44.8 / 100；排名与评分保留该期记录。