Recommending from other people versus recommending from item attributes
There are two genuinely different ways to guess what someone will want next, and almost every recommender system on the internet is either one of them or a blend of both. Collaborative filtering recommends based on patterns of behavior across many users: if people who liked the things you liked also liked some new item, that item gets surfaced to you. Crucially, collaborative filtering never needs to understand what the items actually are. It doesn't know a movie is a horror film or that a product is a pair of running shoes, it only knows which users interacted with which items, and it finds structure in that interaction pattern alone.
Content-based filtering works the opposite way. It recommends based on similarity between item attributes, genre, text description, price range, audio features, whatever can be turned into a feature vector, and matches those attributes against a profile of what a specific user has liked before. It needs no data from other users at all: a content-based system could in principle work perfectly well recommending to a single user in isolation, as long as it has enough signal about that one person's preferences and enough descriptive features about the items. This is the same item-similarity idea covered in the Vector Databases & Embeddings course, applied specifically to preference matching rather than search.