Has anyone else seen this issue? My topics were all working fine on GPT 4o, but now my custom agent uses the Fallback topic incorrectly when it should be using the Thank You topics or other phrase-based topics. 4o had no trouble at all, but 4.1 is struggling.
The Fallback topic for this agent is no different from the system default, except for having removed the link to the Escalate topic as it's not currently in use.
The Thank You topic is simply:
The agent chooses:
Describe what this topic does:
This tool can handle queries like these: Thank you, thanks, ty, thanks so much
(so also basically the default topic.)
I've tried changing the topic wording to a more specific tool-explanatory prompt/description but have had no luck. Fallback takes over everything with 4.1. With 4o being retired on the 27th, and GPT 5 still showing the tool explanation text bug, I'd like to know if there's anything to be done to help fix this before being forced to switch.
Mostly just curious to see if anyone else is experiencing this, or has experienced similar with past version changes.
EDIT:
Sorry folks. It ended up being the issue was due to my mistake configuring the "Conversational Boosting" topic. I had turned it off to prevent web-searched generative answers on my otherwise custom-grounded bot, but that meant then that the Fallback topic was the only topic with "On Unknown Intent" as the trigger. So, anything it was even mildly unsure of would default to Fallback.
The fix was enabling the Conversational Boosting topic again (with only custom knowledge sources used in the topic settings itself, to prevent off-topic conversations),
as well as adding redirects to the "Hello" and "Thank You" topics within Fallback and Convo Boosting based on those key words.
The redirects aren't the most elegant solution but at least it prevents the worst of the "thanks" > "sorry, I didn't understand" shenanigans that were happening.
I do still feel that 4.1 is worse at understanding "Thank you," "Hello", and similar topics, even with directly referenced keywords in the "Agent chooses" section. 4o didn't need the redirects. 4.1 seems overall worse at agent-initiated topics. But, oh well... it happens.
