6 ms·
Cosine similarity alone can't give you good results. For example, if you want to search specific names or acronyms, cosine similarity won't help much. People a
by never_inline 1y ago
Cosine similarity alone can't give you good results. For example, if you want to search specific names or acronyms, cosine similarity won't help much.
People act like embeddings are all you need for search.
Curse of dimensionality also means what you think is the most similar is not necessarily the most similar thing in vector space. See the last Hn discussion on word embeddings for some examples.
- quantadev 1y agoI'm not claiming Cosine Similarity can do things it can't do. I'm claiming it's useful to find related pages, in a very powerful way, and I'm correct.
- never_inline 1y agoYou literally wrote it can be more powerful than Google search.
- quantadev 1y agoBecause Cosine Similarity is that powerful.