A little late to the party here, but text embeddings (at least the ones used in this blog post) generally aren't very good at "searching by vibes": they more compare by overlapping words or look for similar content to the search query.
However, there is a recent paper that actually does try and do this: "Retrieving Texts based on Abstract Descriptions" (Ravfogel et al., 2023) https://arxiv.org/abs/2305.12517.
They give many examples of searching by vibes: "an architect designing a building", "a company which is part of another company", "a book that influenced the development of a genre", etc. etc. Their embeddings apparently facilitate this type of search much better. Would be interesting to retry the offline Wikipedia search from the linked post with this new type of embeddings.
However, there is a recent paper that actually does try and do this: "Retrieving Texts based on Abstract Descriptions" (Ravfogel et al., 2023) https://arxiv.org/abs/2305.12517.
They give many examples of searching by vibes: "an architect designing a building", "a company which is part of another company", "a book that influenced the development of a genre", etc. etc. Their embeddings apparently facilitate this type of search much better. Would be interesting to retry the offline Wikipedia search from the linked post with this new type of embeddings.