Extraction agents: do you window the chat by message count or by time gap?
I'm building a task agent that reads chat messages and pulls out requests and promises ("can you send the quote by tomorrow?", "I'll look into it"). The hard part isn't the model, it's deciding how much context to give it per message.
A builder in a Facebook group pointed out something I hadn't thought through: score intent per message with a short rolling window, and cut that window by a time gap, not a fixed message count. A fixed count breaks the moment someone sends three one-word replies in a row ("ok", "sure", "👍") and pushes the actual request out of the window.
Where I've landed so far:
1. Strictness is about who the message is aimed at (1:1, or you're named in a group), not how formal the wording is. "I'll look into it" from you is a promise.
2. Other people's soft promises go to a separate "waiting" list, so a false positive there costs a glance, not an alert.
Question for anyone building on chat, email or support data: how do you cut the context window, and what gap or size worked for you?
p/raycast
p/kuc-kuc
p/runevr
p/producthunt
p/google-cloud-platform
p/offiai
p/lasso-11
p/many-displays
p/ai-context-flow
p/kodlo
p/reso
p/htmldrag
p/dex-by-exmergo
p/dino-kids
p/jot-2-0
p/beamfy
p/jev
p/supabase
p/nexupira
p/refs
Replies
the ones we lost at my agency started in a meeting and got a yes in email days later
Windowing by message count silently drops early constraints. The agent still reports a clean extraction. What breaks trust is not missing the data, but reporting that everything is fine when it is completely blind to what came before.