I recently make a vector search engine demo[0] that uses neural networks (CLIP) to encode text and image into vector. It runs surprisingly fast consider its hosted on a 2 core virtual cloud running both the text encoder and doing the vector similarity search over a 500k index results for both image and website links.