Hacker News Reader

Comment by RandomWorker | Hacker News Reader

I’ve got this bookmarked and use it to find hyper niche materials on numerical modelling. The stuff it finds on solvers, mesh generation, and optimization methods is so much better than anything I could ever find on Google. Stuff from the 80s and 90s. I’ve found sites written by professionals that I would never find on Google. As someone that doesn’t just take the commercial package of the shelf getting to that knowledge maybe finding Fortran code examples is extremely valuable.

View on HN

Do you have an example of a niche expert page you found useful, which is easier to find on Marginalia than Google?

Search "numerical solver" in any of the major search engines what do you see?

Now do that on Marginalia; you find https://www.scilab.org within the first 10 results, which is an open-source numerical solver software, which get's me to code, which gets me to examples to use.

To be nuanced, I could change my search to "open-source Range-Kutta numerical solver examples" or something better, but why? Give me the weird deeply technical stuff first.

Maybe more a HN example; when I wanted to learn about load balancers, just search "load balancers."

Google; lots of SEO crap (with soooo many ads), youtube videos?, and AWS commercials at the top. No idea, where I'm going.

Marginalia; a linux wiki, nginx (official documentation), a couple blogs by professionals on the topic. yeah, there are some things in here that ain't great.

But If I compare apples to apples the first ten results are just so much better I'd say 8/10 in Marginalia for this example are great to good, while the first 10 (things I could click on) are by companies that don't teach me anything or have articles full of ads.

I don‘t think your examples are very good, these are very generic search terms and if the results are good or bad very much depends on the person searching and what they are looking for.

Underspecified queries ("generic search terms") is actually one of the tricky problems in search. The way Marginalia Search deals with them caters to a particular type of audience and set of usecases. I don't think that's wrong. Seems silly to try to cater to everyone. In that scenario you're more likely to not cater to anyone.

Ah, this helped clarify my doubts. Thanks.

Your "domain" filters seems like a good solution to this?

Yeah. Some of them are a bit rough still, but that's the general idea. Instead of trying to guess what sort of content the user wants, it seems like it makes sense to just give them the option to express that.

The recipe filter is approaching something I'd want to explore further, to be able to provide contextual information outside of the search query.

https://search.marginalia.nu/search?query=cookie+recipe

> particular type of audience and set of usecases

For signed in accounts (which is pretty much is ~3bn Android and/or Chrome users), Google can predict what the user might prefer and yet...

I disagree with this feature, besides the predictable privacy argument, having a search engine transparently serve results according to your tastes makes it really difficult to find things that are new and outside of your existing preferences. It drains the web of serendipity, makes every website feel the same.

Yep. Curating search for an individual makes results worse for a certain class of queries. Opposite of what GP was advocating for, without realising Google already does this.

Exactly: marginalia results are good for geeks; I search for me, not for J Random Consumer.

They're actually perfect examples for this thread.

We've already constrained "what we're looking for" to be "niche expert pages" further up thread. If we're seeing niche expert pages even for generic search results, that's probably a good indication that the search engine behaves the way RandomWorker is describing

It’s result 10 on Kagi.

The quoted phrase finds nothing, unquoted also no scilab, then realized I've made a typo and it's numerical, not numeric, then I get it

Google was ~ top 60, which for such a generic term seems fine, not much scrolling down

kagi will give me scilab as the result #13 and this probably because I raise arxiv and stackoverflow results which will get high.

Reply on news.ycombinator.com