I am not familiar with what actually goes into “neural retrieval”, and it seems to me like semantic search could be implemented in a very straightforward way with no “ai” whatsoever, just by having a database of words that share meanings, and then having less entries in your keyword index. I think the book “Managing Gigabytes” from the 90’s actually mentions this as a possible space saving measure for indexes, and also mentions the possible drawback of finding stuff you weren’t looking for.
You might be able to get away with only collapsing terms in specific cases where there are no other meanings, and thus save space in your index and only help people find things better without the drawback. Can’t think of any words off the top of my head that would work this way but I’m not trying very hard.
As for the term “Vibe Retrieval,” I think I like it. It reminds me of the paper where they discuss how “bullsh•t” is a more appropriate term for what llms produce than “hallucinations,” have you read that one? It’s a good one and I’ll link to it if you’re interested. Maybe “Bullsh•t Retrieval” is more appropriate heh heh
Actually I think there is a database called “wordnet” that you might be able to parse through in this way. You would be looking for words that only have one sense, and are a synonym with another word that only has one sense. Then you could build a list if all the words that are this way, and it would probably be perfectly fine to pretend that they are the same word for the sake of an index.
Now I’m curious as to how many words there are like this. That should be a pretty straightforward programming project.
Well e.g keyword search (or looking something up in the index of a book) does not necessarily retrieve what you want, it just matches words. There's nothing special about semantic similarity in this regard, certainly no parallel with vibe coding.
I think RAG actually refers to something like a chatbot that gets web results behind the scenes and then uses them in its generation context.
I think “semantic search” is a term that I have heard people use regularly for what the OP is talking about.
And then “neural retrieval” which I think only I have used on this page could probably also refer to when they use llms to make ranking decisions for results that were retrieved with a notmal keyword search.
I am not familiar with what actually goes into “neural retrieval”, and it seems to me like semantic search could be implemented in a very straightforward way with no “ai” whatsoever, just by having a database of words that share meanings, and then having less entries in your keyword index. I think the book “Managing Gigabytes” from the 90’s actually mentions this as a possible space saving measure for indexes, and also mentions the possible drawback of finding stuff you weren’t looking for.
You might be able to get away with only collapsing terms in specific cases where there are no other meanings, and thus save space in your index and only help people find things better without the drawback. Can’t think of any words off the top of my head that would work this way but I’m not trying very hard.
As for the term “Vibe Retrieval,” I think I like it. It reminds me of the paper where they discuss how “bullsh•t” is a more appropriate term for what llms produce than “hallucinations,” have you read that one? It’s a good one and I’ll link to it if you’re interested. Maybe “Bullsh•t Retrieval” is more appropriate heh heh
Actually I think there is a database called “wordnet” that you might be able to parse through in this way. You would be looking for words that only have one sense, and are a synonym with another word that only has one sense. Then you could build a list if all the words that are this way, and it would probably be perfectly fine to pretend that they are the same word for the sake of an index.
Now I’m curious as to how many words there are like this. That should be a pretty straightforward programming project.
Yeah, that’s a good point. I’m starting to like this term. It would also be great to read the paper you mentioned—could you share it with me?
Ok here it is: https://link.springer.com/article/10.1007/s10676-024-09775-5 It’s a good read, I think you’ll like it.
Thanks, I like the name lol.
Yes let me find it…
Well e.g keyword search (or looking something up in the index of a book) does not necessarily retrieve what you want, it just matches words. There's nothing special about semantic similarity in this regard, certainly no parallel with vibe coding.
But keyword search is at least a white-box method, it gives you what you want. But semantic similarity is an approximate search and unexplainable.
It's called RAG
I think RAG actually refers to something like a chatbot that gets web results behind the scenes and then uses them in its generation context.
I think “semantic search” is a term that I have heard people use regularly for what the OP is talking about.
And then “neural retrieval” which I think only I have used on this page could probably also refer to when they use llms to make ranking decisions for results that were retrieved with a notmal keyword search.
Yeah I should call it 'Vibe RAG' lol.