Enhance Kagi to Search by Meaning

Pleonasm

The discipline of search seems to be expanding (a) from an old paradigm in which web retrieval is primarily guided by matching keywords in the user’s query (b) to also encompass a new paradigm in which retrieval focuses on matching by meaning and concepts, known as semantic search. The latter is not replacing the former, but is providing an additional capability to maximize the quality of the search experience.

Today, it appears that Kagi only has a rudimentary capability to incorporate meaning into search. To correct this deficiency, Kagi may wish to augment its set of search indices to include one that robustly indexes web content as semantic representations (e.g., Exa).

For example, to experience semantic search in operation and to better understand its contributions to answering queries, construct a query of the form “find academic papers similar to <insert title of paper>” and run the query in Exa Agent using a free account on the Exa Dashboard. The agent will generate a list of highly relevant papers together with an explanation of WHY each paper is similar to one specified, all of which can be exported to a CSV file.

To me, this certainly ‘feels’ like the next generation of search: the tool understands the intent of the query, and matches that meaning against the meaning of web content. This ability to search by meaning is transformative, and unless Kagi enhances its search capabilities accordingly it may gradually become a less useful tool.

mb

As a quick fix maybe it'd be possible to add a custom bang called "find" that opens Kagi Assistant, that should then be able to do the search?

But also, I wonder if Kagi have plans to add Exa as a source.

Pleonasm

The problem of not supporting semantic search is recognized by Kagi. As one employee recently said on Discord, "we have zero semantic understanding of what you are searching for." And, the Kagi CEO understands the underlying technology (embeddings), as evidenced by a paper he authored in 2019. So, hopefully, the company will prioritize the addition of this functionality in order for Kagi to rectify the deficiency and become a leading-edge search product.

P.S.: Personally, I would pay an additional fee for the option of having Kagi search utilize a semantic index.

mb

Pleonasm Well, Kagi uses external indexes, and they must support it.

Pleonasm

When searching in the Mail app in iOS 27, a banner appears stating “Mail can look for more results based on what you mean, not just the words you type.”

Performing a search "based on what you mean" may be a hint that Apple is more extensively incorporating semantic search into iOS 27 and possibly into Apple Intelligence. As semantic search becomes more commonplace and more familiar to consumers, perhaps Kagi will need to adopt the technology in order to remain a competitive search provider?

mb

Pleonasm I wonder if it'd be possible to interpret the queries and reconstruct them, so that they work with traditional search engines. If you want PDFs then it's filetype:pdf. But, what if you want articles and documentation about PDFs, e.g. as a software engineer? Or you want a browser tool for merging and splitting PDFs?

user12345

mb interpret the queries and reconstruct them, so that they work with traditional search engines

I think that's a good idea. Prompt: "Hey, could you rephrase my search request, so that it better expresses what I want."

Pleonasm

user12345 I think that's a good idea. Prompt: "Hey, could you rephrase my search request, so that it better expresses what I want."

A web search initiated by an AI Assistant (e.g., Kagi Research Assistant) automatically rewrites a user's query to better define the question, and also breaks that question into subquestions for additional clarity. But it still conducts the search utilizing a traditional web index that is primarily focused on keywords.

In contrast, semantic search relies on a web index that encapsulates the meaning of each webpage, thereby allowing the user's intent to be matched to the meaning of web content.

mb

Pleonasm what indexes are available for that?

Pleonasm

mb Pleonasm what indexes are available for that?

The Exa search index is built to encode the semantic meaning of web content (as described here).

To see the impact of this technology, run a query in Exa Agent using a free account on the Exa Dashboard of the form "find webpages similar to <URL>". The agent will generate a list of highly related webpages together with an explanation of WHY each webpage is similar to one specified (all of which feels like a "surprise & delight" experience). In contrast, Kagi Search cannot find webpages that are similar to an exemplar webpage, because the tool has no inherent and underlying concept of content similarity (beyond rudimentary overlap of keywords).

For some very simple and specific queries (such as searching for a product model number), the traditional keyword search may be superior. But, for more thoughtful and complex queries, a semantic search may add value in a way that is not currently possible with Kagi. Ideally, Kagi will add the semantic component to its search product to enable both approaches.

user12345

I don't know if you already do this, but here's my Idea:
Find fitting search results by consulting an AI in the backend on your server about the search request of the user. You could prompt this AI, which is hidden to the user, like this: "Hello, this is Kagi Search, an Google alternative. The user is searching for "XYZ". What search results would you give him?".
And then take the links the AI gives you and display it to the user.

I'd like to know if you're already doing this, and if not, what you think of the idea.

Better search results, both web and image content.
It would be a bit like an automatic version of "Quick Answer", but only compiling links.
One could even go so far and tell the user why the AI recommended a specific result to him.

Duplicake

You can't feed your whole link index to an LLM to rank, it's simply not possible. If it did a search with kagi and then reranked them with AI I think the results wouldn't really change that much since the main issue is finding what links to put in the top 5 not how to rank them. On top of this it would be extremely expensive and quite slow

user12345

Duplicake the main issue is finding what links to put in the top 5 not how to rank them

I agree. I don't want the LLM to rank, but to find links. Is that possible? I mean, I can ask ChatGPT and Claude: "Hey, could you suggest a light summer dish?" and it works.

Duplicake

user12345 that's because it searches using the exa engine and it gives the top 10 links to the LLM which it then reads and generates an answer, it's not ranking itself.

Pleonasm

Firefox has now partnered with Exa to power AI search in its browser.

The CEO of Exa noted, “Our philosophy is actually we want to raise all ships.” As another 'ship' in the 'ocean of search,' perhaps Kagi could actively explore establishing a similar partnership with Exa and thereby ‘raise’ its level of search excellence?

foxberg34

Semantic search sucks.
Instead of searching, the model has to try to guess what kind of result you're looking for, and then guess what the result looks like, and then find the results. So instead of getting a mix of anything with the word "shoe", the model guesses that you want to compare prices for shoes near your location, and returns 20 shopping sites. Which is why all Google results end up looking the same.
I would much prefer if anyone could give me a dumb keyword search.

Pleonasm

@foxberg34, searching for "shoes" might not be the best test case for semantic search, since the query is so very basic.

However, since we're on the topic of shoes, you may be interested in the experience of one individual using the entity discovery "websets" feature of Exa Search....

Let me tell you about a real problem I had, and how Websets swooped in like a superhero. I was trying to find companies in Turkey that sell affordable shoes. Sounds easy enough, right? Wrong! Normal search engines? Forget about it. I was getting bombarded with websites selling expensive designer shoes, random shoe blogs, and online stores that shipped from… everywhere except Turkey. It was a digital shoe-finding nightmare!...
Then, I remembered Websets. I thought, "Let's see if this AI thing can actually find me some affordable shoes in Turkey." So, I typed in: "Companies that sell shoes in Turkey cheap." And BAM! Like finding a hidden shoe store with a giant "SALE" sign, results popped up. Not just any websites, but a clear list of companies, their website links, and even little checkmarks saying "Yep, they sell shoes!" and "Yep, they operate in Turkey!".
But it wasn't just a list of names and links, oh no! Websets went deeper. For each company, it showed me why it thought they fit my search....

🙂