This was a very meandering project, and trying to corral it into some sort of coherent narrative was a bit of an undertaking on its own. Hopefully it makes some sense.
Hi Viktor! Really cool write-up, thanks! Uruky is already using the nsfw param, but set to 0 or 1, and I see in your example this looks like a new value option (2) that's "better" than 1? How "safe" is it to implement it as the value to send when someone wants SFW results?
Wouldn't work very well, in that you'd get awful recall.
The way the filter is implemented, it runs after the query has been executed. I'd have to run it at document processing time, code in a pseudo-keyword for the label, and then add that to the query.
It's doable, but I question whether the juice is worth the squeeze.
Looks like a cool search engine! Hadn't heard about it before.
But the search page says "Simple technology, no AI". With this change, that is no longer true though, is it? Of course the definition of "AI" is extremely vague. Once upon a time, A-star search was considered AI after all.
The 'No AI' statement is about gen AI, which is I think what most people think of when you say AI.
But sure if I was looking for government or research funding, then for sure this would be AI. Not just AI, but the literal state of the art AI. Dario wakes in a flop sweat every night, terrified of my breakneck advances in single hidden layer classifiers that are probably at least 30% sentient. It would be so much AI I can't even hold all this AI.
Does marginalia_nu not use embedding models as part of search? I guess I assumed it would. If you have embeddings anyway, decision trees on the embedding vector (e.g. catboost) tend to work pretty well. Fine-tuning modernbert works even better but probably won't meet the criteria of "really fast and run well on CPUs". That said, the approach described in the article seems to work well enough and obviously provides extremely cheap inference
Have you seen many examples of websites labeling themselves, perhaps using rating meta tags ()? Self-labeling seems valuable in some ways, but I don't think I've seen it catch on.
16 comments
nsfwparam, but set to0or1, and I see in your example this looks like a new value option (2) that's "better" than1? How "safe" is it to implement it as the value to send when someone wants SFW results?1 filters 'harmful' sites per the UT1 blacklists
2 is 1 + the new NSFW filter.
The new filter works pretty good in my assessment. It's not infallible, but it gives significantly cleaner results.
And if you do find queries it fails to sanitize, I'd love to hear about them.
So I can make sure I know what sites to stay away from, of course
The way the filter is implemented, it runs after the query has been executed. I'd have to run it at document processing time, code in a pseudo-keyword for the label, and then add that to the query.
It's doable, but I question whether the juice is worth the squeeze.
But the search page says "Simple technology, no AI". With this change, that is no longer true though, is it? Of course the definition of "AI" is extremely vague. Once upon a time, A-star search was considered AI after all.
But sure if I was looking for government or research funding, then for sure this would be AI. Not just AI, but the literal state of the art AI. Dario wakes in a flop sweat every night, terrified of my breakneck advances in single hidden layer classifiers that are probably at least 30% sentient. It would be so much AI I can't even hold all this AI.