I already have a couple of user that have been complaining that the bot stops mid-sentence, like it’s writing an email sequence and all of a sudden stops. Sometimes when you prompt it to continue, it continues, but that’s not how it should work.
My advice has been to ask the bot for shorter answers so it takes less tokens, but still they keep getting their answers cut off. We have seen it mostly in two of our people, and the chat does it several times in the same conversation.
Thanks for flagging this, Kirsten!
Before we pull deeper logs, please go into the Config tab for that pickaxe, open the Output section, and bump the max output length so the model is allowed to finish longer replies. When that slider stays at the default it will cut the agent off midsentence (same thing the community called out over here: Chat GPT 4o responses are getting cut off fairly quickly - #4 by luna_support).
After you save the higher limit and refresh the portal, let me know if chats still stop early. If they do, send me the pickaxe/portal URL plus timestamps + member emails and I’ll grab the logs immediately.
This is not a config issue. Seems to be happening even with high amount of output tokens. For example, set output to 800 and answer output is only 68 tokens.
Got it, thanks for testing the higher limit. To grab the logs we need the specifics from one of the failed runs (pickaxe link, portal link, member email + timestamp or conversation ID). Once I have that I can pull the completion traces and see whether the model is throwing use-limit errors, content policy stops, or something upstream that is cutting replies at ~70 tokens. Feel free to email info@pickaxeproject.com if you’d rather not post the workspace details publicly and I’ll keep this thread updated with what we find.
Hi folks, just following up to reiterate what Luna said - we’d be happy to look into this if you can send over this info!