Hacker News Reader

Comment by foobarbecue | Hacker News Reader

Doing planetary science, I used to be in the habit of searching for image IDs on google. Usually particular orbital images of the moon which had been mentioned in a paper. There are of course specific archives to search for this stuff, but google used to give surprisingly good results. Now I get literally zero results when I search an image ID. What changed? Did they remove their index of the Planetary Data System? Did they remove the indexes of planetary science papers?

View on HN

Part numbers, error codes, and similar "high entropy" terms also seem to suffer the same fate.

Even worse than giving no results is giving you page after page of "results" containing slightly different numbers/codes.

No worse than searching for words and having pages that don’t contain them.

Same with phone numbers. Most of what is returned has suspicious sites in .ru, .pl, et al.

I don’t think they removed the results from index, but I believe they started searching for vectors instead of text matches. Your ids are nowhere in vector space.

Is this why back in the day if you typed in 2 works it would show results with those 2 works and how the majority of time just the first?

When I'm looking for a technical solution I combine words. But Google gives me the middle finger, returns results for the first one which is like 'Windows Service' only to show me a generic website which has a download link for every KB number out there.

For the query “where is Microsoft headquarters” they want to return results with phrases like “Microsoft is located at …”

This makes non technical question-answering easier, to the chagrin of every techie.

I think your problem is a mix between a lot of things, but mostly it seems to be SEO farming.

Shouldn't it be possible to work around this by wrapping the query in double quotes? Google promises to do an exact match search when wrapping in quotes.

So I'd imagine the OP's query should work when the image id is wrapped in double quotes. Otherwise, it'd mean that it's not in the index anymore.

Except Google lies.

It's like trying to write a Gmail filter with the "=" and then a word wrapped in quotes. You would THINK that would be a strict X = Y... but it's not.

At some point a few years ago the Gmail team made the "=" operator a "fuzzy equals", so for example it matches on parts of a word, even if you explicitly say you want to filter a word with a space at the end because it's "close enough".

Drives me fucking mental - I have an email filter that instantly archives every email from Jira except those that have my user name in the body of the notification (meaning someone has tagged me).

In Outlook I simple filter for "my_username " (with a space at the end), and it's able to tell the difference between that and my email address. In Gmail, "my_username " somehow matches both "my_username" and "my_username@email.com"... meaning EVERY FUCKING EMAIL I GET MATCHES THE CRITERIA.

Yep, I tried with double quotes, both times.

I studied literature, and used to be able to find the most obscure tip-of-my-tongue references by describing fragments I remembered of a text or paper.

Same result: Now it's nothing, or unrelated SEO.

Try to use yandex instead, it is order of magnitude ahead of google when it comes to image search. Edit: sorry I was convinced you are searching by image, not image id.

It sounds like your suggestion came about by mis-reading the parent comment, but I will offer my 2 cents: yandex.com's english results for hard to find, technical terms are infinitely better than google's.

as someone who remembers How Google Used to Work, Yandex feels like Old Google. It indexes the hell out of long tails, will find exact, specific terms and phrases, and 100% respects generic search operators such as exclude "-".

do I look for news articles on it? hell no. do I find the unfindable (on google, and increasingly, on duckduckgo)? yes, I do.

Yandex is also usually quite a bit better at finding pirated things in my experience

My suspect is BERT and semantic search. Before it most word embedding models were monolingual and non contextual, so they were not useful for search. So Google had to use inverse indexes and an algorithm fine tuned for years.

With BERT they can have a model to embed all the search queries for 99.99% languages and works sort of fine for 99% of the queries, so they are fine with that as well.

It would have been sort of easy detecting that a query works better with the previous index than the new one, but it seems that they decided against that. So we lost a lot of usefulness, like phone number search.

Can you give an example query and the page it should return?

Sure. Searching M1104273380R should return https://wms.lroc.asu.edu/lroc/view_lroc/LRO-L-LROC-2-EDR-V1.... and this paper: https://doi.org/10.1038/nature19829 .

Interestingly, when I tried this just now, I did get the paper as a result. I could swear I tried this same search yesterday and had zero results.

I've heard that Google will process queries that don't seem to get results for the searcher and they improve if you come back to it. I never test this because I've usually given up (wasn't important) or moved on to other ways to find the information.

Not sure about Google but DDG changes the results when I come back to them. So, for example, if you search for something, click on one of the links to open it and then go back to the search results, there will be a couple of new ones added.

Not sure if some are removed also but it definitely changes.

In a ironic twist, it now returns this HN page.

Google has multiple search clusters, and indecies are not fully in sync, since they are updated asynchronously. So often the difference is the cluster your query is landing on.

Assuming you know the domain that usually gives you the result you’re looking for, I would just search <IMAGE ID> site:name.domain

"There are of course specific archives to search for this stuff, but google used to give surprisingly good results."

Reply on news.ycombinator.com