Comment on Why I Lost Faith in Kagi

hersh@literature.cafe ⁨7⁩ ⁨months⁩ ago

I’ve been using Kagi for a while, so I’ll post a few quick thoughts I had after reading the article, linked blog, and mastodon thread.

The one thing in the blog post I strongly disagree with is her statement that the summarizer is “the same old AI bullshit”. I think they just assumed that without actually testing it. The summarizer is fantastic, and is a great example of the right way to use LLMs. Its output comes entirely from the URL or file you specify. It does not hallucinate. You can ask it follow-up questions about the document, and again, its replies are limited in scope to what’s actually in that document. If you ask it something out of scope it’ll tell you that it can’t find that information in the document. This is great because it’s using the LLM for what LLMs are actually good for — complex language parsing — and not for what they’re bad for, like reasoning or information storage/retrieval. It’s actually quite difficult to misuse the summarizer. It’s straightforward and effective. This is Kagi’s killer feature, IMO.

I can’t speak as highly of its search-integrated AI features like FastGPT. They mostly take information from the contents of the first few search results, but they also seem to “fill in the blanks” more than I’d like. Sometimes it gives me information that is simply not in the sources that it cites. It’s not as bad as using ChatGPT (which hallucinates all day every day, and if it ever tries to cite source is hallucinates those, too) but it needs improvement.

That said, Kagi doesn’t shove the AI down your throat like you might think reading the blog post. These are separate features that need to be explicitly activated. Plain searches don’t return results from the LLMs. If you want that, you need to explicitly go to the assistant or trigger the “quick answer” feature on the search results page.

Anyway, back to the summarizer, here is an example of it in action. I gave it the URL of the Mastodon post. I think this is an excellent result. I also asked it an unrelated followup question so you can see that it limits itself to the information in the URL. It will not make shit up!

The summarizer lets me download conversations in markdown form, so I’ll just paste it right in here so you can see.

hackers.town/@lori/112255132348…

Exploring https://hackers.town/@lori/112255132348604770

Assistant:
Key points:

[1] lori (@lori@hackers.town)

Query:
What’s the capital of North Dakota?

Assistant:
The knowledge provided does not contain any information about the capital of North Dakota. The context is focused on an email exchange between the author and the CEO of Kagi search engine regarding criticism of the company.

source
Sort:hotnewtop