I care a lot about the correctness of information and the speed at which I can retrieve such information. LLM and the quick answer feature in kagi search, greatly improve the speed at which I can get information, but I am always worried about the correctness. Sometimes the LLM will use an untrustworthy source and sometimes the LLM will misunderstand the source, or it will leave out important context. This is why I think the source links that go directly to the relevant location on the source page is awesome! However, going to the source like that is an extra hurdle that I think a lot of users don't take. That is the reason for my proposal:
Since LLM are probabalistic, they can not be trusted to even copy text correctly from a source.
This is why I think it should be possible to have LLM responses that contains verified verbatim quotes from online sources. These quotes should be:
- Easy to distinguish (perhaps a different background color)
- Clicking on the quote should take me to the exact copy of the same text on the source website.
These two properties combined make it possible to trust the quoted text without going trough the effort of verifying the source text myself.
This is the only way I can trust any information that comes from a LLM model.
Perhaps it would also help to instruct the LLM models to do less paraphrasing and more quoting to make use of this new feature. Or give the LLM model some method to request a verbatim quotation so that it can guarantee a perfect match.
For example, if the user goes to Kagi search with a query such as "Who is the president of France?", then the LLM could respond with something like
According to wikipedia: "Emmanuel Jean-Michel Frédéric Macron[b] (born 21 December 1977) is a French politician who has served as President of France since 2017." [wikipedia]
The quote would be highlighted and is an exact copy of the source text. Clicking on the quote would bring the user to the same snippet on wikipedia. They can see that the quote is from wikipedia without clicking on the quote, again without the LLM being able to tamper with the cited source.
This gives the user 100% confidence that they are not reading a hallucination.
