If you are ranking documents based on KNN-style algos against document embeddings, how do you deal with more intentional search queries that have a narrow intent to retrieve a specific type of content?
If I browse around and save content to your app, I might want to retrieve this exact piece of content and not any other one that may be semantically close to the query. My experience with vector-based systems (and let's be honest, we are really talking about a recommendation system rather than a search algo here) is that they are hard to combine in a way that is conducive to efficient information retrieval.
You search for a "dog" and you get "cats" because they are similar, but maybe I really want "dog" and not "cat". I find that tremendously inconvenient when apps force you into semantic search and don't let you choose. Sure that might increase the recall, but still please let me choose.
If I browse around and save content to your app, I might want to retrieve this exact piece of content and not any other one that may be semantically close to the query. My experience with vector-based systems (and let's be honest, we are really talking about a recommendation system rather than a search algo here) is that they are hard to combine in a way that is conducive to efficient information retrieval.
You search for a "dog" and you get "cats" because they are similar, but maybe I really want "dog" and not "cat". I find that tremendously inconvenient when apps force you into semantic search and don't let you choose. Sure that might increase the recall, but still please let me choose.