Does that mean you'd return other docs if they share just one word?

The idea of tfidf is that it gives you a vector (maybe combined with pca or a random dimensionality reduction) that you can use just like an Ada embedding. But you still need vector search.