Clarification on the usecases for Kagi Assistant

RoxyRoxyRoxy

Bodine The Assistant has a lot of models (with new ones added very often!), letting you use and test out some of the best LLMs across providers. We also have a beta Assistant called Ki, that adds image generation and deep research (leveraging Kagi's own search results and user customisations). We're also currently looking into ways to implement a memory function similar to ChatGPT. Let me know if you have any more questions or things you're curious about with Assistant, I'm happy to answer them!

Bodine

RoxyRoxyRoxy It's good to hear that image gen and deep research are in the works. But I'm still a little concerned that the Kagi assistant will constantly be in a game of 'catch-up' to other LLM apps. What I would like to see is features incorporated that other assistants can't easily provide. Something like considering users web page ranking adjustments when searching for sources. Or maybe the ability to mark certain websites as 'preferred source' when browsing. I haven't thought about it for too long, but I'm sure there are many possibilities from making the search engine and assistant more intertwined– things that ChatGPT or Claude can't do.

Maybe another way of looking at it is that Kagi should focus more on features like the Universal Summarizer and 'quick answer.' Trying to build its own unique space and feature set with LLMs would be much more useful than just trying to create a Kagi version chat interfaces– as the chat interfaces provided by the the respective model makers will always be better unless Kagi introduces features that they can't.

Edit: Also you mention that Kagi has a model selector. This is an example of a good unique feature! But on it's own it doesn't provide much after the user has decided which is their favorite imo. Perhaps there are some possibilities here like having two models work together to generate a response, either by passing info back and forth or by having each generate their own and some way of displaying the best one. Again, I haven't thought much about it, but it's an interesting avenue that other LLM providers couldn't necessarily do.

madhatter

How does one sign up to try Ki?

RoxyRoxyRoxy

Bodine

Something like considering users web page ranking adjustments when searching for sources.

We already do this, your full domain customisations are carried over from Kagi Search into Assistant : )
You can also use Lenses with Assistant to further customise what sites to search. I believe there was some discussion here about that sort of dual-model approach, not sure if we ever explored heavily into that, I'll check!

EDIT: Now that I think about it, would you mind making a new post for the dual-model Assistant suggestion? This helps to keep things a bit more organised 🙂

Flint_6

I tried ChatGPT, Perplexity and Gemini extensively along Kagi Assistant.

I like Perplexity for its various models. But it's crowded with features like newsfeeds. It has Spaces that work really well to create custom projects with attached files and links, but I found links are not to be trusted. I don't see when Perplexity is really going to a link and indexing into the context and several times it had no clue there was something attached. So weird bugs there.
Furthermore, I hate the layout. Chats just get listed with the FULL question in the sidebar and it makes it so hard to find conversations if you don't diligently rename them. And within the chat, Perplexitys layout makes it super hard distinguishing where my questions are and what the answers were. In a long discussion you're playing a game of hide and seek. Perplexity has an annoying voice mode on mobile that is basically useless to me to hold any discussions.

ChatGPT has a great layout, great projects, a great chat mode. It's one of the best apps of them all that I used. Super clear in its design. Only downside is: You can only use ChatGPT, which is notoriously happy and thinks every crap I produce is the best in the world. Golden! What an idea!

I really like Kagi and I come back to Kagi Assistant often to use other AIs. It's fast, it's clean, it's strength is in its simplicity. I like you always see what AI gave the answer. An info you have to click to see in Perplexity. The answers are plain in the design, which makes it easy to read. I'd like a better visual different between what I input and the AI responded. ChatGPT has a great chat interface for that. As if I'm typing in WhatsApp. Easy to see what was written by me.

Features I miss in Kagi: ImageGen (apparently being worked on). I'd like Audio Transcriptions from audio files. I'd like a better voice input mode so I don't have to type long instructions into the input field - the voice mode is useless sadly. if I only pause for half a second to think what to say, the assistant stops listening. no conversation or longer dictation possible. Super annoying.

I'm sure at some point I'll switch to it full time.

kirkmc

Flint_6 As for audio transcriptions, that probably uses a lot of tokens, which will lead users to hit their ceiling pretty quickly. I'm a journalist, and I use transcriptions regularly, and it's no surprise that good online services are not cheap (or have hard limits).

Related: if you're on a Mac, there's an app called MacWhisper Pro that can do transcriptions locally. It's almost as good as Otter; so good that I'm cancelling my Otter annual subscription when it runs out.

kirkmc

I use Perplexity. This is all I see. You can turn stuff off in the settings.

ChrisP

kirkmc Oh I'll give this a try.

What I'm doing currently, I have a SetApp subscription and included is a WhisperTranscribe app. Normally the subscription to this app would have the price tag of an Otter and the like.

But you can hook it up to the OpenAI API. I paid $10 once weeks ago and for a 1hr long podcast it only charges me a few cents. So every now and then I just need to top up my API account and that's it.

Would be handy though not having so many moving parts and just do transcriptions where I do all my AI work.

Thanks for the hint!

ChrisP

kirkmc Holy moly MacWhisper is a great tool and I haven't had this on my radar. Thank you so much kirkmc for pointing me to it.

kirkmc

ChrisP I've been using it since it was released, at first plugging it into the OpenAI API. It started out okay, and has made big improvements, notable adding speaker detection a few months ago. It still has a way to go, as segments aren't always split at the right spot, but for my use, transcribing interviews which I then use for articles, it's great.

And I like that everything is done on my Mac, not fed to a company to train their LLM.