Quick answer giving incorrect answer

@john79 thank you for providing feedback! Could you please help me to understand if the right article was part of the top three search results? 

Hi @tetiana12 ,

Yes, the top result in my screenshot “Prisma: How do I reset my password for Spectra AV?” would've been the right article. That article doesn't contain much content in itself as the answer is in an attached PDF. 

@john79 Zendesk search, including the generative search feature, doesn't search within article attachments.

Understood @tetiana12 . I'd be less concerned of the wrong article being suggested if the quick answer didn't change the content of the article within the summary.

If the quick answer said, “To reset your password for Spectra OX” - it'd be clear to the end user that it's not the right answer. Can we get any more detail on why it changed the opening line to “To reset your password for Spectra AV” when Spectra AV isn't referenced anywhere in the article?

Unfortunately, it appears that this is a case of an OpenAI hallucination. We understand your concerns and are currently researching ways to prevent such occurrences in the future.

I ran into a similar result, too. 

Do you have information on the number of required thumbs down before the model learns? I expect some incorrect information on occasion. I'm more concerned about the effort required for it to improve based on the negative feedback.

Related: Has anyone added some add'l guidance for users, encouraging them to use the thumb up / down?

Hi @chris92 ,

I posted the same question on the Generative search for help center (EAP) article. I tested this and gave an incorrect answer a thumbs down a total of 11 times as both a licensed agent and a test end user. I continue to get the same incorrect answer on subsequent searches. 

We had a developer embed text above the quick answer, example below, and were going to add text along the lines of “Help improve the accuracy of quick answers by marking results as helpful / not helpful.” However, we've held off until we get clarity from Zendesk about how the model learns.

Thanks @john79 - I'd like to get a good feel for this. While it seems slow to react to your feedback, I also wondered on how would it be handled if the feedback was incorrect enough times – training it incorrectly. We will likely need to change our process where today we let anyone create an account, before we open this up for feedback. We had generated a ton of community spam until we required each new post be moderated, because the spammers were creating new accounts as fast as we shut them down.

Hi @tetiana12 , copying my comment from the help center article to the community:

Can you provide any insight on how impactful the thumbs up/down feedback is? How many thumbs down does it take to stop suggesting a quick answer? Is the impact weighted differently depending on the reason they chose for negative feedback? 

Also, is there a way for us to review the free text feedback a user enters when they downvote?

Hi everyone! Zendesk generative search is powered by OpenAI. They cannot use our customers' data to train their models. We review your feedback manually and adjust the prompts we use for OpenAI.

@tetiana12 - Thanks for helping us understand the process. I work with Open AI / ChatGPT a significant amount, and I'm very familiar with refining prompts.  I can only imagine this is a new process for your org to figure out as well. I recognize we're in early access, so I'm sure some of these details are still being refined.

Can you tell us a bit more about your process? What would cause you to adjust a prompt? 1 feedback 10 feedback? It depends/still working on finding the right process?

Is there a way we can help you verify when a prompt is adjusted?