Comment by leggomylibro | Hacker News Reader

Comment by leggomylibro | Hacker News Reader

I think that Adafruit uses them, too.

What I notice with their results is that they start strong, but fall off very rapidly into partial gibberish.

For example, say I want to make a battery-powered project. I search "lipo." The results start out promising; charging boards, including a few options tailored to specific popular boards. But by the time we get to the end of page 1, only every other result or so is actually relevant. Useful results like batteries and connectors are interspersed with random microcontrollers, LEDs, motors, etc.

Still, it works and you can always refine by category if that sort of thing really bothers you. It seems like a pretty solid solution, and the issues I'm mentioning are probably caused by the implementation, which I'll bet uses the full text description of entries to weight things towards not excluding a potentially-relevant result.

Ideally you'll get everything you need in the first page anyways but generally speaking if you have complex part numbers or product series, it's harder to get right. I think that machine learning is essential to improve relevance over time without having to resort to manual tweaks.

Adafruit does have it tough; they try to bridge sites like Amazon and Digikey by having a very broad selection of products specific to electronics, but they don't actually have many parts in each of those categories. They focus more on supporting what they have with documentation and software and whatnot.

But sites like Digikey/Mouser/etc have searching down. If I want a capacitor, first I pick what kind (electrolytic, ceramic, tantalum, etc,) and then I am presented with dozens of menus representing specific attributes that I care about. Capacitance, temperature coefficient, size/pakaging, manufacturer- whatever you could possibly want to select on, you can.

Sometimes I wish that other digital distribution platforms would take inspiration from that 'catalog' model. Discovery is difficult, these days.

Sounds more like a design decision - don't exclude documents, where there is a low probability of relevance. For various reasons it might be better to have more results with some irrelevant ones, instead of 7 highly relevant ones.

They're slightly prioritizing recall over precision: https://en.wikipedia.org/wiki/Precision_and_recall

Reply on news.ycombinator.com