Separate the information from moral advice

Kagibeh

I recently looked up whether the word XY can be used as a despicable term for small dogs.

The AI, after answering my question, proceeded to tell me that it's not nice to use despicable terms for small dogs, and that all dogs should be treated with respect and appreciation.

Main problem in this case isn't the lecturing on good/bad manners (we've been there before), but that it was unclear if it was part of the actual reply.

Part of the actual reply means: This information was found in the sources cited and it's an organic part of the information relayed by the AI. But the AI provided no reference to its claim that all dogs should be treated with respect and appreciation.

Not part of the actual reply means: The advice was added artificially and reflects the private opinion of the programmers or the search engine's legal team. But in this case, I feel it should be placed in a box or somehow separated from the actual reply, just like advertising should not be mixed with information to make clear which is which.

silvenga

I've also been noticing this, across all LLM's. As Kagi uses the public models, I don't think Kagi can really do anything (as this isn't something that can easily be fixed with a better system prompt).

All the common models have this training ingrained to be almost annoyingly polite and "politically correct". There are quite a few attempts to weigh base models against this, although, it's hard to remove such aggressive training completely. It's also assumed that a lot of the training data comes from things like Reddit (filtered by sentiment) and generally "clean" sources (e.g. text books, PG novels, etc.).

You could try to customize the prompt used (the "Custom Instructions"), and see if that helps - perhaps by instructing to output more terse responses?

Value7609

silvenga I've built an internal documentation chat bot at work (using RAG), and I've included a phrase in the system something like: "Only use the information given, otherwise tell that you can't answer this question", and this worked really well it only answered the questions given the documentation snippets and otherwise it would say that the documentation does not include this information.

This is not the exact situation as the problem with quick answers (another sentence would be needed), but I'm fairly sure that this behaviour could be at least reduced (definitely not "solved" thats for sure) through a good system prompt.

Kagibeh

If this can be solved as easily as adding a standard phrase to the system prompt, then it should be just as easy to make it a Kagi switch, no? For the switch, the phrase would then just be added in the background.