I agree with you that I'd love to know more about how it works. The citations to real supporting websites are generally very good (though I've seen a few mistakes).
I think it's a great complement to human lead search.
If only Kagi had a Shopping feature, then it would be a complete solution for us.
Why?
presumably they want a shopping feature so that they can shop, and otherwise find Kagi has all of the features they want
Right... why? Is it really too much after talking to the assistant and finding the item you want to purchase, to then click the provided link to the product page and click buy?
I do not want a click-to-buy button, I want kind of lens that is specialised for shopping within specific regions (the current kagi lenses allow only 10 sites which is not really enough for shopping).
That is quite literally what !research does.
But agreed on local search - I tend to use Google maps for anything local. Otherwise for something I'd buy online, ¯\_(ツ)_/¯
In contrast, I may start with a page in a website and start going around looking for related products etc. The assistant is not gonna start doing that (I think at least), it just gonna get the page provided by the search results. For this reason, what search results it is provided with is important.
I am also (mostly) buying online, but for certain reasons I want to buy from domestic websites rather than ones based in other countries, if possible. Kagi has not been great at providing great results for that.
But, regular models and harnesses work okay for this too. You just have to specify “web search”, very important.
I use pi and librechat and I am really limited by the lack of proper web connector. Testing Tavily right now, Kagi is next.
I wonder why I see so few comments about it on HN, does everybody have this figured out already? Am I a dummy? Am I severely overstating how important of a frature web access and search is for LLM use?
The one caveat is opus has an absolutely awful voice. The way it writes is annoying and verbose. Sonnet actually works better for this type of thing.
If I'm paying for it through Kagi, I could at least trust their incentives are aligned with mine instead of whoever is paying them the most.
On the one hand, it's made me a believer in AI as it relates to web searching.
On the other hand, the only reason I need AI to search the web is that the web is now a impenetrable forest of near unreadable, SEO-stuffed, spammy articles.
Unfortunately, that is nothing new. A big part of the reason why Google gained popularity, and maintained it for so long, was its resistance to SEO stuffed, spammy articles. The spammers continued escalating things, and now it pretty much takes AI to avoid their garbage.
I suspect that product and service recommendations will be a near impossible problem to solve. Any time there is a potential for financial gain, someone will figure out how to game the system. Even if it could be resolved, the queries would have to focus on objective criteria ("best" is too subjective).
Is it actually, really? Even late 90s most stuff was garbage, and been like that for as long as I've been on the internet, or maybe I've just always been jaded. But the ratio of "garbage to non-garbage" seems constant for as long as I can remember, multiple decades at this point. The tools for finding the signal through the noise have changed over time though.
It doesn’t change the fact that they literally impacted the trajectory of how the internet developed in a very obvious, concrete way.
Part of it might just be that it’s not rammed down your throat anywhere, there’s no sparkle or pop up to “help me write” - google and Microsoft have destroyed their interfaces with ads for AI features. Kagi only does anything if you know what it is and want to use it.
You can get it on at least iOS as well as its own app. You can select from bunch of China-models, I use Kim pretty much always. You can also do a system prompt and bunch of the same stuff you can with Claude etc. But it lies to me hardly ever. I don’t know how they did that.
It will say that nothing in the provided search context has your answer which seems like a simple system prompt but somehow other models from bigger companies have a much harder time saying "I don't know" by default.
I do remember them asking about shipping information at some point though.
I'm finding though ChatGPT with sol 5.6 is as good or better, so I'm using that primarily these days. Can also do actions in browsers on your behalf, so you can have it book your whole vacation, etc. Pretty rad.
There is a lot that agents can help with wrt search, but they also have a lot of blind spots. They work well many times, but then there are times where I find stuff that multiple llms missed. Imo this is gonna work as in many other domains, as in the more different tools one is proficient with (searching with llms, searching traditionally etc) the more successfully one solves problems.
ChatGPT is trained on all of Reddit and internet forums where most of the useful info is. It just automates the drudgery.
EDIT: I see I have $5 of "pre-paid API credit". Is this a one-time thing or is that included with my plan?
Edit: this is a one-time promotional credit. One must pay for credits thereafter, in addition to their subscription.
I tested on my local LLM chat via Open WebUI, 2 requests from that cost $0.02 (so $4.98 left), and after my monthly Kagi renewal an hour ago, it did not reset, I am still on $4.98.
You are probably better off using Kagi Assistant anyways, as you get $10 there on the $10 Professional plan.
For local LLM web searches, to save on cost I will stick with the Brave Search API which gives you $5 free every month.
which model?
So this has nothing to do with the Kagi AI assistant then... it's just the model that's perceived as "good". My point is, there are cheaper ways to get access to the model, without a Kagi subscription...
Kagi Assistant defaults to the last-used model unless you got there from a Quick Answer. Several times I forgot to change models, and interrupted K2.7 Code, but after a failing to notice the model a few times and getting usable results, I now usually just leave it alone. Unless I want a quicker answer and doubt a "Code" model will grok me.
I'm here to admit I was wrong, it's incredibly useful. It misses the mark frequently, like any LLM, but it's well annotated with footnotes that link to the original material.
I use it all the time.
I also love how easy it is to use or not - just add a question mark to the end of your query to use the AI, or leave it off for traditional search.
With that UX it doesn't feel like it's being shoved down my throat.
Creating an account just to hate on a company trying to shake up a monolithic and user-predatory product space is just sad.
Meanwhile google translate is probably served on the cheapest model they could muster.
https://github.com/isene/ClaudeCode/blob/main/agents/web-res...
This is honestly also one of the reasons why I don't subscribe to Kagi anymore. My 'Google-Fu', combined with whatever personalization is happening on my account, already keeps Google's search results in a similarly useful state to Kagi's, so Google and ChatGPT already cover my needs.
Paying for Kagi Ultimate feels like a privacy upgrade but a quality sidegrade that isn't worth $270 a year.
Kagi also has a Small Web feature which is meant to surface exactly that sort of thing. It's not amazing, but I appreciate it existing. I think you can get a "random small web page" kind of thing even without a subscription here: https://kagi.com/smallweb/
I think there is a lot of space for improvement on "searching skills" though, sometimes the search strings they use are not great.
But if you're using the starter plan, which is 300 searches for $5, then you definitely do pay for the assistant. The API cost would be $3.60.
I like this too, but it isn't as bulletproof as it sounds. I've run into two instances in the past week where it confidently gave me wrong information, because it ran into a search result that had wrong information, and it trusted that information.
You could perhaps say that's fine; if you were doing the search yourself, you'd get the same results and might trust the wrong information too. But the "might" there is... oh boy, I'm gonna say it... load-bearing: someone is much more likely to believe the confident LLM response than if they were looking directly at the primary source, along with search results with other sources that might contradict it.
The assistant is still good and I also use it regularly. It’s also nice to not have to leave the terminal sometimes.
Nobody speaks about their AI Assistant, but it is really good. They somehow harnessed it so that it mainly searches info first and sticks to the verifiable data. I prefer it to Claude or any other, because it actually answers the question I need without fluff or “Great question! Here’s some plausible nonsense you can trip over instead of doing actual research!”.