Allow Editors to Provide Feedback and Correct Eddy AI Answers
V
Vijay Sakthivel
Currently, feedback on Eddy AI answers can only be submitted by readers through likes, dislikes, and comments. Editors reviewing Eddy AI Analytics cannot mark an answer as incorrect or provide the correct answer when no reader feedback has been submitted.
It would be helpful to allow editors to review Eddy AI answers from the portal, mark inaccurate or irrelevant responses, and provide the expected answer or additional context. This would help teams refine Eddy AI responses and improve answer accuracy, especially when the source documentation is correct but Eddy AI does not identify or connect the appropriate context.
Log In
S
Similar Coyote
Interesting but not sure about this. Isn't the correct workflow to update the content and test the question again? That way the content improves over time. Where we've had cases of incorrect answers, it's normally traceable to conflicting information or incorrect information in the source content. But willing to learn more here.
G
Gossamer Lamprey
Similar Coyote: We've definitely been improving the source content and retesting questions, and Accurate content is always the goal. The challenge we're seeing is that there are still cases where the AI selects a response that is technically related but not the correct context, even after the documentation has been clarified.
What I'm suggesting isn't a replacement for maintaining good documentation. It's an additional feedback loop where knowledge base editors can indicate, "This response was not the one we wanted the AI to provide," and optionally identify the preferred response or correct knowledge path. That kind of editorial feedback could help the AI learn from cases where multiple articles contain similar terminology, but only one actually answers the user's question.
For example, several products may use the same feature names, while one product has a different workflow documented in its own article. Even when that distinction is clearly documented, the Eddy still favors the more commonly viewed or broadly related content because of the shared terminology.
Today, unless an end user submits feedback, editors have no way to flag that response as incorrect or provide guidance from the knowledge management side without replaying the end user's conversation repeatedly.
The goal isn't to bypass content improvements—it's to give content owners a way to continuously improve the AI's answers in situations where the documentation is already accurate, but the AI's response selection still needs refinement.
V
Vijay Sakthivel
Raised on behalf of Gossamer Lamprey