It's neat but I've found the value kinda variable. It seems heavily influenced by whatever the first few hits are for a query based on your question, so if it's the kind of question that can be answered with a simple search it works well. But of course those are the kinds of questions where you need it the least.
I find myself much more often using their "Quick Answer" feature, which shows a brief LLM answer above the results themselves. Makes it easier to see where it's getting things from and whether I need to try the question a different way.
The quick answer (ending searches in a question mark) also seems pretty resilient to hallucinations. It prefers telling you that something wasn't mentioned in the top search results over just making something up
There is one more aspect of Kagi assistant that I don't see discussed here. I'd love to support some "mass tipping jar" service and/or "self hosted agent" that would benefit site owners after my AI actions spammed them.
You can simply just pass it a direct link to some data, if you feel it's more appropriate. It works amazingly well in their multistep Ki model.
It's capable of creating code that does analysis I asked for with moderate amount of issues (mostly things like it used the wrong file extracted from .zip, but it's math/code is in general correct). Scraps url/downloads files/unarchives/analyses content/creates code to produce result I asked/runs that code.
This is the first time I really see AI helping me do tasks I would otherwise not attempt due to lack of experience or time.
I find myself much more often using their "Quick Answer" feature, which shows a brief LLM answer above the results themselves. Makes it easier to see where it's getting things from and whether I need to try the question a different way.