Aaron, thanks again. I think that will be beneficial. I am no information specialist and more of a clinician and my work (in workers compensation board) involve a lot of literature searches; systematic reviews etc. I am pretty sure that I am not a lone in this kind of work situation. I also hope that you can spread your writing wider so that many more people can learn it. So, thanks again; I've been following your post for a while and learnt quite a bit from your "musing"
Hi Aaron, thanks for this, it's very useful. You mentioned that Smart Search results are reproducible when filtering for Boolean (which is like not using Smart Search?). Would you expect to be able to reproduce results when filtering by semantic search? Once concern I have about AI is its lack of transparency and what I think might be the lack of reproducibility of results when using AI. Do you think that's a fair concern?
I goofed a little. What I mean is the Boolean search component of SmartSearch is exactly the same as doing a normal WOS Boolean search. As such this component is interpretable/explainable, just like a normal boolean search.
Semantic Search is now known to be using vector embeddings. This is by definition black box and not interpretable/explainable (or in your terms transparent).
Reproducible means something different. A search is reproducible if you run the exact same query, you will get the same result (assuming no index change). The Boolean search component part of SmartSearch is reproducible as well.
I am not sure about the semantic search, often it isn't reproducible if it uses vector embeddings, but so far, I've tested it is. Always gives same set of Semantic Search results in my testing.
Aaron, thank you for keep on writing on the issue of AI and searching. It is quite confusing to me.
Yes it's confusing. Glad you liked my posts. I'll try to cover more of such "semantic search"
Aaron, thanks again. I think that will be beneficial. I am no information specialist and more of a clinician and my work (in workers compensation board) involve a lot of literature searches; systematic reviews etc. I am pretty sure that I am not a lone in this kind of work situation. I also hope that you can spread your writing wider so that many more people can learn it. So, thanks again; I've been following your post for a while and learnt quite a bit from your "musing"
This is a brief simplified intro to embeddings in retrieval. https://aarontay.substack.com/p/why-embedding-vector-search-is-probably
Hi Aaron, thanks for this, it's very useful. You mentioned that Smart Search results are reproducible when filtering for Boolean (which is like not using Smart Search?). Would you expect to be able to reproduce results when filtering by semantic search? Once concern I have about AI is its lack of transparency and what I think might be the lack of reproducibility of results when using AI. Do you think that's a fair concern?
I goofed a little. What I mean is the Boolean search component of SmartSearch is exactly the same as doing a normal WOS Boolean search. As such this component is interpretable/explainable, just like a normal boolean search.
Semantic Search is now known to be using vector embeddings. This is by definition black box and not interpretable/explainable (or in your terms transparent).
Reproducible means something different. A search is reproducible if you run the exact same query, you will get the same result (assuming no index change). The Boolean search component part of SmartSearch is reproducible as well.
I am not sure about the semantic search, often it isn't reproducible if it uses vector embeddings, but so far, I've tested it is. Always gives same set of Semantic Search results in my testing.