[mattbaker]

I’ve been using Kagi full time and I like it a lot. It’s been worth the price.

I expected to like lenses and favoring/blocking specific domains. What I didn’t expect was how much their “Quick Answer” would change how I search.

I’ve been “AI hesitant”, in general the chance that an LLM will hallucinate makes these kinds of tools more trouble than they’re worth for me personally. In Kagi’s case, though, the individual facts it states in the quick answers have citations linking to the site it drew that information from.

Here’s what I’ve found:

- it’s been accurate most of the time, but not 100% (as expected)

- citations are pretty accurate most of the time

- every so often the citation links to a page that seemingly doesn’t back the claim in the quick answer

Unsurprisingly, I don’t trust the AI generated quick answer in isolation, what it does do is let me scan a few paragraphs, find the one that answers my question most specifically, and visit the sites it links to as citations for that piece of the answer. This saves me the time of clicking through the top $N results and scanning each page to find the one that seems to answer my query most directly. It’s like a layer on top of the page rank.

I remember using Google the first time and being impressed how the top answers were so much more relevant than Yahoo, it was a huge time saver. Now I find myself wondering if the “quick answer” citations will prove to be a similar jump in accelerating my ability to find the right web page.

It also makes me wonder if their own page rank algorithm could incorporate the quick answer output as an input to a site’s rank? That would be an interesting experiment!

[freedomben]

Same experience with quick answer. I originally thought of it as a gimmick, but it's really changed the way I search now. The summaries are good, and the citations are amazing. I (nearly) always click through to the cited sites to decide how much I trust them, but the vast majority of the time the summary is spot on and extremely helpful.

On a side note, as a software dev I was curious about how it would "answer" queries that weren't questions, so I've tried feeding it queries like "linux distro" or other things that are most certainly not questions. About half the time it does a pretty good job at answering! For the really hard ones, it seems to just morph it into "what is <query>?" so nothing real complicated, but I did find it interesting.

[joshuakogut]

What happens when it crawls some LLM generated text on a website, using that as a citation?

[jabroni_salad]

It gives you in-line citations so you can see which parts of the text might come from a subgrade source.