Extend feedback options in Interactive tester
under review
D
Daffodil Reptile
Currently, giving a thumbs down on an AI Agent response offers limited feedback. To improve learning and response quality, we propose adding predefined feedback categories (e.g. factually incorrect, incomplete, or poor style) and an optional text field. This makes it easier for users to clarify what went wrong without writing full new instructions.
Similar to ChatGPT, this system would help identify patterns in user dissatisfaction and offer the AI Agent clearer input to improve future answers. It would also save time, as users wouldn’t need to update domain knowledge or instructions each time a small tweak is needed.
This post was marked as
under review
D
Daffodil Reptile
With a thumbs down, you now have to add a completely new instruction.
Perhaps with defined feedback categories, you can train an AI agent faster because he now learns from the mistakes made?
We would love to hear about your experiences here.
I have additions or comments to many of the answers. Instead of constantly making substantive changes to the domain knowledge or adding new instructions, it might be more efficient to be able to provide the reason or explanation for each answer. This way, the AI agent knows better what the adjustment is based on and where to look specifically to improve/optimize the answer.
These will be the following 3 main pillars:
- Answer is wrong/factually incorrect
- The answer is incomplete/additional information needed
- I don't like the style (for example, the writing style, the length of the response, or the tone: too direct, too cumbersome, or not appropriate to the situation)
> In certain contexts or conversations, it is desirable to give a short, to-the-point answer, while in other cases, a more comprehensive and explanatory answer is needed. Instead of predetermining this with fixed instructions, as an employee, you can indicate what you prefer by using the thumb icon)
With ChatGPT, you have both defined feedback categories and the ability to write customized feedback. Only then does it become a bit more of an instruction. Perhaps good to have 3 to 5 in advance as they occur most often.
S
Sufficient Deer
Daffodil Reptile I totally agree, we're experiencing the same thing.
O
Orange Mastodon
Daffodil Reptile I agree with this too!
Fleur Nouwens
Hello Daffodil Reptile! I have a few more questions for you:
- What specific types of feedback options would be most useful for your needs?
- How often do you find the need to provide detailed feedback beyond a simple thumbs up or down?
- Would you prefer predefined feedback categories or the ability to write custom feedback?