Kagi Assistant
blog.kagi.com
blog.kagi.com
I've been using Kagi for a while (almost two years now!) and it's been nothing but excellent!
Lenses are very useful (Reddit lens is on every second search), and I personally really like the AI features they are working on.
The new more advanced assistant which is able to do searches, which can also be constrained to lenses, and lets you pick an arbitrary model, is excellent, and basically means I don't need a chatgpt/claude subscription, as Kagi covers it very well.
All in all, great product which I'm happy to pay for.
https://kagi.com/search?q=remarkable+pro+site%3Areddit.com
It includes Reddit results from less than 24 hours ago.
If I search for that exact Reddit post using the lens it isn't there. https://kagi.com/search?q=ReMarkable+Paper+Pro+hands-on+revi...
If you google “best DSLR camera reddit” it’s much less valuable if the results are 5 years old, even if they’re LLM free - the cameras on the market may very well have changed in that timeframe.
I skimmed down and if this was about Kagu indexing being out of date then sorry I just realized that
Perhaps this is because I'm in Europe and it's faster in the US? A search request to Kagi seems to take around two seconds for me (shows as ~1s in the Kagi UI), it just feels really unpleasant compared to Google, I'm used to firing of a couple searches with different wording / terms and go through results quickly, feels like I'm being held back.
Maybe I'm spoiled, but if I'm paying for search I would really like it to be at least on par with Google, the search result quality seems ok from what I can tell, lenses don't really make sense to me, they seem to filter out too many results I would have liked to actually see, but the customization like adjusting the rating of individual websites is fantastic.
If they can manage to bring the speed to match that of Google, I'd be happy to pay for it I think.
I'm not slamming you, or your experience or preferences. That said, I find very little difference between 2 seconds and one second or less than that.
Search taking a small fragment of time just isn't a big deal and I search sometimes many times per day.
What is the gain for you that makes 2 seconds an exception to using the product?
Just curious. Peace, live well.
Because it’s such a common action for me, it feels like such a strong regression to go 2/3 times slower than before.
I still remember when google was giving relevant results in page 2. Now it's pretty much useless for me, and the fast search makes me think they are throwing away tons of potentially good stuff just to make it fast (and place more ads and rubbish scam/ai sites).
Google fully pivoted from prestigious tech-company to lame ad-company and that's a shame really. Business as usual.
Back in the day, IBM did a usability study. An application requiring the user to specify operations and fields for data input was setup two ways:
One way was manual, lots of clicks, and or input sequences. Each one did a specific thing quickly.
The other way was highly automated and the user was only required to click a few times. The tasks were the same. This way had more flow, fewer discrete commands, more functionality woven together.
To their surprise the users felt they were more productive with the software they clicked more, despite the automated version being less work and the workflow more efficient.
I bet the effective reduction in your search is minor, but the delays do accumulate and demand attention.
Given that, a small change to your flow may well change things!
What you should do is rapidly input your first queries and then as they appear, drill down on those, and when that appears, start to eliminate dead query windows, or drop fresh queries into them.
What you prefer, to use a car analogy, is one that corners like no other. Then you find yourself in one that lacks corner cases.
So you maximize your time in lane, straight road, batching the corner driving and flooring it on the straights.
I used to experience a similar thing running a browser on IRIX vd NT. The NT browser was quicker to respond where the IRIX one would delay a little and then just render it all quick
I just started working with a few windows. Changed everything. I would be typing in new queries while one I waited for was about to render.
It was a change from rapid fire to a more batch mode. Soon, I rarely had to think and my flow was fast all around.
I put this shared experience here in the hope you may be inclined to try different things.
https://glinden.blogspot.com/2006/11/marissa-mayer-at-web-20...
The other day I was using someone else's computer and used Google, and my goodness, the results were just awful and ... bloated?
1. they simple pull ads from database on the first page instead of actually searching for what you need
2. They load marginally relevant answers instead of doing a better search for what you need
3. Google is multibillion company and can afford faster servers
With Google you waste a lot more than 2 second scrolling down past the ads trying to find answers, assuming you can even find it, and doing another search(es) if you don’t, wasting even more time.
But do what you want.
If so: that's your root cause.
You may consider reaching out to them about what you are seeing, it might be something they could investigate and resolve if they know about it.
I can confirm that the search is randomly slower than it should be. Sometimes hanging indefinitely and I have to refresh.
The added second or so to search is manageable but noticeable for sure.
To be fair/clear, it’s often unnoticeable but when it does hang it drives me crazy.
Interesting, as I see a lot of other comments from Europeans saying they don’t notice this delay, maybe it’s Germany-specific?
What does bother me is the occasional failure to load anything at all, solvable with F5 key. But, being able to rank pages, or rewrite reddit.com to old.reddit.com more than compensates for this.
Status Connected to: EUROPE-WEST2 Network latency: 47ms
But I don't really care about AI assistants. If these AI integrations will improve the bog-standard search experience (I type something in the search bar, get a list of results, and click the one I think will give me what I want), that's great. But if not, this is just noise to me.
What I want most of all from a search engine is to be "internet plumbing" and mainly stay out of my way, and with Kagi I always had the feeling they would suddenly remove/change/add things because of some strong opinion their founder holds. In which case I don't want to be $108 committed for a year.
I love Kagi and can’t recommend it enough. I wish I could just give them my api key for this instead of paying several different service providers for the same ai access to the same models. This is getting expensive.
Any chance you can expand on this?
I've used DDG for years and find the results great, but maybe we just query for different things. Is there any specific topics that you find DDG is bad at but Kagi is good at?
That being said, if you personally are happy with your DDG results, then I'd say you should probably stick with it. Kagi _might_ be better for you than DDG, but if you aren't actively dissatisfied (the way I was with Google and everything else), then the room for improvement is smaller. Might be worth using their free tier (which is 100 queries per month I think) just to test out some side by side searches though.
100 in total, it seems
With Kagi, I haven't had that. It finds what I need and surfaces things that Google didn't. On the one or two occasions I've had trouble and jumped to Google out of desperation, I also didn't find what I was looking for there... so I guess an answer didn't exist.
For sites I don't like, I can block them from my results, or rank things higher/lower. I can make my own custom !bangs and shortcuts, instead of being stuck with what I'm given. While I haven't used Google since they launched their AI features, I can say I have really enjoyed Kagi's AI responses to questions, and that each line of the response is referenced so I can check out the source (not sure if they all do this or not... I can only compare it to ChatGPT, which does not).
I've stopped using Google due to their captcha being so frequent and frustrating. I don't want to spend 60+ seconds answering 6 captchas purely because I use Firefox with privacy settings/extensions.
I usually search in a private tab, enhanced tracking protection set to strict, uBlock with most default filters enabled.
Oh and a "Google Search Fixer" extension because the make the UI much worse for Firefox. Maybe it's one of my extensions, but DDG never traps me in captcha hell.
Why do you ask people instead of trying it and finding out if you like it or not? They have a trial of 100 free searches.
This is a little bit like asking a coffee drinker "explain why I should drink coffee". It's better to try for yourself and see if you like it.
Yeah, you're right, weird of me to do that.
Being able to personalize your own search is truly the killer feature, though, in a couple of ways.
The first is as you point out: being able to "edit your own algorithm" is really nice. I don't have to try to "train" Google's algorithm to show the results I want, and it's very easy to say "I never want to see this site in my results again". I'm still shocked Google doesn't have that feature even as some kind of client-side Javascript.
The second is Lenses. It's so obvious in hindsight that a singular algorithm is insufficient for search. Nobody wants or needs their searches for porn to impact their searches for technical documentation, or vice versa. There are more nuanced examples, but that's the most obvious (also, I don't think Kagi indexes NSFW content or at least I haven't seen any).
https://addons.mozilla.org/en-US/firefox/addon/ublacklist
First two extensions I always install on a new computer are uBlockOrigin and uBlackList.
Honestly it’s awesome I use a software at work so I upvote that site and block the site of the competing software for example so I get the correct help guide. I don’t understand why google doesn’t offer many of these features.
I imagine the intersect between paid Kagi users and paid LLM users is pretty high, and many people probably don't want to double-dip on LLM spending.
https://github.com/raycast/extensions/tree/99c7c7c4fa02afba9...
It essentially summarizes the top search results for you, leaning in on a strength of LLMs (summarizing) while reducing its greatest weakness (hallucinations).
Kagi has AI generated stubs for some of the search results, probably originating from some of the search indexes they pay for.
You pay a different rate per model (OpenRouter shows the pricing transparently). You load your account with credits. I use it daily (undoubtedly far more than the average user) and loaded 50$ with credits five months ago, but I still have over 1/2 of it left.
I think it is hard to believe that Kagi would be any cheaper and have no rate limits.
This means that it's more expensive than calling OpenAI directly, even though they have the same price per token.
If you want to use precisely one API, paying directly for that API is cheaper. However, that's only true with closed-source providers. Anyone can host a server running llama 3.1 that OpenRouter could (in theory) use, bringing price competition to model cost. Closed-source models have a monopoly and can set their price wherever they want.
I'm okay with spending an extra 2$ every six months to access the APIs of any model I want.
If it has a downside, I'd say its a little more involved to get setup, lots of docker containers, etc then a more batteries-included approach like Jan.
https://openrouter.ai/docs/limits
...haven't had issues with them, but they are there
And if you cared and loaded your credits, you can do 200 req/s (maybe also a surge limit? docs are unclear)
If you just want text chat with different models it's great.
Here's their blog post: https://www.spreadprivacy.com/ai-chat/
It's since been updated to include newer models: https://duckduckgo.com/duckduckgo-help-pages/aichat/
Whenever I try this with Google, I can't seem to get it to work though. According to [0], it is not an official search operator anymore.
0 - https://support.google.com/websearch/answer/2466433?hl=en
Just to make sure we're not accidentally giving Google a break, Google+ ended in 2019, if they cared for search at all they would have brought back this feature.
That’s what eventually made me cancel. Loved it for what it does give but found myself rather annoyed about having to switch to Google to get actual usable results on the significant amount of location based queries. Never got into the habit of using bangs unfortunately.
FWIW DDG is flawed in nearly the exact same way.
One feature here that I think competitors lack is that the LLM's view of search results can be constrained by Kagi's search "lens" [0] that let you exclude various categories of results.
I use Kagi but haven't dug into lenses, anyone have experience ?
I'm currently trying to write python script interfacing with outlook's mailterm interface (win32com.client) and it's annoying. I wonder if I can restrict search results to a particular domain so it only pulls from microsoft docs...
Note: I work part-time at Kagi (was a Kagi user before that), not doing search stuff.
Yeah, you can set up an allowlist of domains in the lens configuration. I have a "Reddit" lens that essentially just adds "site:reddit.com" to the searches
While I am still happy with the search, I find the value of the assistant no longer worth it. At first, I thought I would be able to replace other providers with the wide range of offered models, but now I find myself often going to distinct providers. Kagi seems to have introduced a harsh and generalized input token limit for prompts (detached from what the underlying model can handle), which makes it almost impossible to work with code or documents. I think the VW example in the announcement makes it clear that the assistant is only intended to respond to prompts of a few sentences—searching by sentences instead of keywords. I don't see input for documents or things like code execution coming here, so the difference of $15 monthly can get me fancier features.
Mixing the output of the model with the limited content of approximately five websites (= Assistant with internet access), on average, is most of the time counterproductive for me. The sites mentioned in the sources of the output are often not sites I would usually visit or get information from (they usually don't correspond to the top results on Kagi web search). I have used search engines for decades now and have built a sort of index/pattern/feeling in my head for which site I am going to visit, which AI can not match for me.
The lenses are nice but are also pretty specialized. Most of the time I find myself just doing a broad query, which in most cases will already satisfy me.
I will definitely continue to use Kagi for search. There is no comparable search experience out there, and it makes searching definitely more effective.
This field is moving very quickly and since there will always be a lag in getting new features/models into the "wrappers" I'd rather get it from the source.
One random thing I wish Kagi could do is offer a way to promote the "official" website to the top of the search results. I have various up/down rankings applied but when I search for "DataDog" I want their official website at the top of the list, not under SO/Reddit/etc posts. If I'm searching for a problem then yes, I want those sites higher but it's slightly frustrating to have to scroll down 3-5 results to find the main website for a product/service/company. I feel like they should be able to differentiate between "search about X product" vs "X product".
Funny you mention Datadog, I specifically struggle to keep their blog/promotion/etc official results out of the way when digging around for documentation and troubleshooting. If I’m trying to find a solution to something weird with Datadog I’m often facing the opposite of your problem and wish Datadog’s site would get out of the way.
By already having a traditional search engine this puts Kagi at a big advantage compared to someone like Perplexity, or even Claude and OpenAI who I think are all cobbling together solutions on top of Bing's API.
Want to search for open source projects that implement some algorithm? Create a Github/Gitlab lens. Want to ask it questions only about some framework? Add it’s domain to a lens.
Being able to innovate on the search side while everyone else loses a lot of many training LLMs feels like a good space to be in at the moment.
But maybe my search topics aren’t niche enough.
If you can focus the chatbot on a specific docs website with a lens though that problem might go away.
Don't want pinterest in your search results? Block it:
For the first while, the search results were inarguably better than any alternative. I was thrilled and was recommending Kagi to anyone who'd listen. In the last six months or so, things have gone downhill. More often, my results ignore some of my search terms and I have to try to "trick" Kagi into matching them all. More often, I get only one or two pages of results when other engines give me notably more relevant hits. And more often, my search hangs or fails to return at all.
I don't care if they noodle with AI or make T-shirts as long as the search is great. What I've experienced is the search getting worse while they noodle with AI and make T-shirts, though. The two may be entirely unrelated but from an outside perspective, it feels like their core offering is suffering because of these other-than-search undertakings.
So far, I'm sticking with it, but my enthusiasm has definitely diminished. A couple of my friends have canceled their subscriptions in disappointment and I've started to consider cancelling mine, as well.
Mistral and GPT - Create the same example of a flip card
Gemini - Creates glowing neon text
and Claude - Produces a pulsing dot, that enlarges and shrinks and radiates a fading white shadow. That's cool.
Stop launching new products (browser, summarizer, gpt, assistant) while your core product is still behind the competition in many areas.
Whether or not Kagi can achieve more than the "search alone is hard enough" point however is fair - though i've been happy so far.
To me, the best part of the AI additions is that it can (almost instantaneously) summarize information from the several top hits of a search. This is subtly but importantly different from having the LLM spit out an answer based on it's knowledge base, and also is able to quickly and easily cite it's sources! Extremely useful to me.
Also, AI aside, Searx has been around for many years as a very promising metasearch, even self-hosted engine, alas still little traction. Great to have all results, including re-ranked Google and Reddit in one place.
> Kagi Assistant has the ability to use Kagi Search to source the highest quality information meaning that its responses are grounded in the most up-to-date factual information while disregarding most “spam” and “made for advertising” sites with our unique ranking algorithm and user search personalizations on top.
They're literally taking on competition here.
The people I know who like to dictate into their phones and who have OpenAI's iOS app tend to open ChatGPT to "search" before they open Google or Ask Siri now.
They're going to go to one thing first, and this puts Kagi as an option.
Apple's alleged integration with OpenAI is presumably rolling in to Sherlock this though.
Brave Search has offered an AI summarizer and assistant for a long time now. Bing with their OpenAI-powered Copilot. Google with the improved Bard/Gemini more recently. Amazon with the perhaps Anthropic-based Q for Business.
I think the end user is growing to expect the AI-augmented experience from all knowledge lookups. Feedback loop queries have become so natural to me, I’ve been finding it awkward to ask only one search query without a narrowing follow-up query, having the former discarded - kinda no longer adequate, particularly given the SEO-optimized flip side full of junk.
IDK, I've been very happy with it. Just the ability to consistently pin/block domains is a massive upgrade over Google.
I think of it this way: there's often not a single page (or even small handful of pages) that answer a query. The LLM features answer the question with text that links to the pages, rather than answering my question with pages that might contain pieces of answers.
Care to explain what exactly do you mean by this?
I might end up disabling web search until they've tweaked it a little more. I really like kagi search so I feel pretty confident that they'll get it, but I don't think it's ready to replace the brand-name chatbots.
> No markdown blocks for code makes assistants really tough to use.
Could you clarify if you are referring to the markdown from your prompt not rendering? Or the response from the model not rendering markdown for code blocks?
The former was an intentional choice for now to simplify release; it is on our roadmap. The latter would be a bug - in which case, you can email the thread URL to zac@kagi.com and I'd be happy to take a closer look.
In my daily work as an MD it's become my reflexive go-to for looking up answers to specific to general, easy to complex clinical questions. I use it far more frequently than UpToDate (which is no less than the holy book of medicine), more than PubMed/Google Scholar searches, and definitely more than a basic web search (Google, only b/c it's a hassle to log in to Kagi every session at work).
Maybe 1 time out of 10 it won't give a correct or meaningful answer (in which case my prompt needs to be refined, or is just not suited for this kind of tool). But apart from that it will give me exactly what I need, because it uses Kagi search to source its answers. Kagi search does a decent job bringing to the top relevant journal articles (which in turn may mention other articles, adding indirectly to the trove of sources FastGPT pulls its final answer from). It shows the 5 search results it referenced at the bottom of the page, so more often than not if I don't get my answer in the direct summary, I have very relevant sources to read through.
I also don't think you need a Kagi account to use it.
The thing stopping me currently from trying this or Claude is I rely on the Opt+Space shortcut with the ChatGPT mac app.
Are there any other options for a native mac app with integration as good as the ChatGPT app?
https://github.com/raycast/extensions/tree/99c7c7c4fa02afba9...
Is it still legal do do something with a computer without involving "ai"?
You only get the quick summary if you ask for it (i.e. you're happy to take the risk). Another handy use, is web summarize, which looks at the specific webpage, lowering the risk even further.
If you don't call these features it's straight up clean search with no ads and it's fast with good results.
This mid-thread editing feature sounds really useful, I'm curious how does it work when you switch between models in the middle of a conversation?
Like, say I start with a general search question, then halfway through I want to switch to a coding model to ask something like, "Can you create a Python dictionary of the top 10 longest city names in the UK and their populations?"
Does the context carry over smoothly, or would I need to rephrase things when switching models? Wondering how it handles tasks that require different kinds of expertise without losing track of the flow.
among other things, it randomly says I have no connectivity in an awful modal, I click to search and the keyboard doesn't open, I hit "enter" to search and nothing happens (I need to tap on the search suggestion). pretty bad so far. will try it again in a couple of weeks to see if there is any progress...
What it is though, is fast, available on all my devices, constantly upgraded, and integrated with their already excellent search engine.
When I see these sorts of announcements and read some of the comments here, it makes me worry that bad customers cause enshittification and I hope kagi stays true to their human-friendly web search product.
Update: I see now that they say it's not available for free users. Need to pay $25/month. Not sure why, they can offer it for free users with the cheaper models like they do now to generate a "quick answer". I'm not going to pay to try it out.
For many queries side-by-side with Startpage it delivered the same results word-for-word (sure you get a few sponsored links top-3 of Startpage but its no big deal to scroll past those).
For other things, it was just plain annoying, e.g. "newest $type restaurants in $large_city" half the results on the first page were from 10 years ago (e.g. dated 2014). I mean FFS I put the word "newest" in there !
They seem to have a habit of interespersing very weird Facebook links randomly in the middle of a list of results. For example I was searching for something related to a specific Prometheus function (which I explicitly named in the query, alongside the word prometheus) and Kagi insisted on interspersing the technical results with random links to Facebook pages of companies selling "girlie dresses for proms".
I approached Kagi with an open mind, but having used up the 100 free searches nothing made me say "just shut up and take my money".
For some of us it is. If your search engine's revenue model is based on advertising to its users, their relationship is fundamentally adversarial. This affects all of their decisions, in ways that are sometimes hard to identify. Witness the slow decline of google search result ads.
If users are the direct source of revenue, then everyone's interests are aligned.
Also, I, and many like me, value a lack of ads much more highly than you do. Which is fine.
It's hard to beat.
For example, I don't want to log into or provide payment information to my search provider, because I don't want to voluntarily provide personal information. Compared to Kagi, I can use traditional ad-based search in a relatively anonymous way.
The other big part of it (for me at least) is seeing more obscure websites in my results. I have had Kagi for a year now, and it has saved me more money than I've spent on it by making it easier to find specific products at lesser-known shops. These lesser-known shops often have really great sales because they are trying to compete against the big names, and Google pretty much only shows me the big names.
I get that you don't mind, but for me, I would find that intolerable.
> PyLLMs is a minimal Python library to connect to various Language Models (LLMs) with a built-in model performance benchmark.
If you're going to use this feature, always check the citations.
Imagine for a second a world where instead of publishing directly to your own website (home), you publish into the LLM's knowledge base. Nobody comes to you for information, they come to the LLM. Nobody cares about you, or the work you put in. Your labour is just a means to their end. To some extent, you could argue that Wikipedia works in this way, but it really doesn't. The work you put in is reproduced verbatim, and the work is collaborative. You get the joy of seeing your writing being used to help other people in a very direct sort of way, as opposed to being aggregated, misinterpreted and generally warped by a non-intelligent LLM.
In other words, you cannot possibly expect others to want to work in a sweat shop, toiling away to provide you with instant gratification. We must leave room for human expression.
If you are publishing information for free, you don't care how people access it.
If you are publishing just information without selling anything and want to make a living on it, you should paywall it or get with a publisher.
Just added to announcement FAQ
how do i _ in python !chat