[User Feedback] spam detection should filter messages uniformly without taking into account capitalization

[ryan953]

Original Message that was marked as spam: https://demo.sentry.io/feedback/?feedbackSlug=react%3A5513579551&mailbox=ignored&project=5808623&referrer=feedback_list_page&statsPeriod=90d Message: THis was slow and then I cant buy

Another message, not marked as spam: https://demo.sentry.io/feedback/?feedbackSlug=react%3A5513619151&mailbox=unresolved&project=5808623&referrer=feedback_list_page&statsPeriod=90d Message: This was slow and then I cant buy

We think that the mixed casing "THis" vs "This" messed up the filter. Both messages should not be marked as spam

We do not take into account the name for spam detection

[drguthals]

I don't know how the model works , but similar potential things to explore if we are looking at adjusting this:

• Common mis-spellings (isn't automatically spam)
• Apps that want to leverage UF for other spoken languages

[aliu39]

We do not take into account the name for spam detection

I think it makes sense for name/email to not be taken into account, since it could confuse the model. Since we allow anonymity/empty values for these 2 fields, we don't do any validation on them

[aliu39]

There's 2 ways to solve this:

1. lower case the whole feedback message before passing it to the LLM. Surefire fix.
2. Include verbal instructions to "ignore casing and make a best guess at misspellings" in the prompt. Could be more robust, but need to validate results

Both should have no problem solving the casing issue. I'll try out different instructions for the misspellings. Something to think about is how badly a word can be misspelled before being interpreted as junk