OpenAI content filter exception handling

[jhicks-nfg]

Encountering something interesting regarding OpenAI content management policy. We deal with a potentially hazardous substance, and it appears that OpenAI's filtering periodically regards some requests as related to "violence." Sometimes this results in the system responding with "sorry, I cannot assist you with that..." but other times it fails to record the conversation item and dumps that raw error event right into the user's chat experience. (see screenshot)

1. Is the error being dumped into the chat response expected, or is that a bug in this sample app? Inconsistent UX implies bug. Perhaps a bug in the chat history recording process when there is a content issue?
2. Is there a way to track if the OpenAI filtering is constantly preventing responses to someone's chats? Perhaps due to what we would expect to be normal behavior, but perhaps due to a problem user?
3. Since we do deal with a dangerous substance, is there a means to adjust OpenAI filtering rules so the system realizes discussion of such substances do not have violent intent within our context? Not seeing any env vars that relate to that...

Appreciate any thoughts!