Kagi Assistant is searching through way too few, and sometimes the wrong sites when using Web Access

Mar

Kagi Assistant often only has 5-6 sources when using Web Access.
That makes the answer heavily reliant on these few sources and if I have a pretty general question while the results have very specific information (happens a lot to me) the AI will talk about this specific case, but not my actual general question or it might not contain the relevant details I need, while a broader search would.

I would like the answers to contain more information and have better reliance on actual sources.
This could just be solved by introducing more sources I think.
For example when I use perplexity it often searches for 30-40 sources before giving me an answer.
Also, this could maybe be achieved by searching not once, but multiple times, if the AI finds it doesn't have the proper results yet or should search something else, it creates a new query that's more fitting to what was actually searched for.

howie

Web searching is noted as one of the most expensive parts of the entire Kagi Assistant process, and is a part of how some people reach acceptable usage limits very quickly - I imagine searching for 5-6x more sources than currently could be unsustainable

silvenga

Yeah, I've found assistant (both OpenAI and LaMDA being my choices) tends to "hyper-fixate" on the first few sources - which sometimes is good, but on more complex topics, it quickly causes hallucinates.

That said, again, depending on the query, assistant just ignores the search/sources and only uses it's memory (seemingly from picking the wrong sources that were actually unrelated).

RoxyRoxyRoxy

Maybe it could be an option to enable “extended search”? Toggling that on would enable Assistant to search through many more sites. Off by default of course. Not sure how Kagi could balance usage of it wrt cost, maybe once Assistant is on the Pro tier extended search could be Ultimate exclusive? I definitely don’t need more than 5 sites for 95 percent of Assistant usage but it sure would come in handy on that last 5 percent.

mixchiev

It would be nice if kagi made a custom R1-based assistant or sth that could be more search-focused (like the deep-reseach features of google/openai). Ime getting LLMs work well with search is a huge challenge and feeding the context with more search results can also mean feeding the context with more irrelevant results, and what LLMs are definitely not known for is solving the frame problem. But now that we have a good and relatively cheap reasoning model, maybe it is worth a try.

Rehan

We actually have something in the works that would address this. We call it Ki, it we have it in beta right now, but it is able to do a lot more searching and retrieval compared to the current assistant.