An embedding is a list of numbers that places a piece of text somewhere in a space where distance means similarity of meaning. That sentence is the whole idea, and once it is genuinely intuitive rather than memorised, a great deal of modern AI stops being mysterious.
The course builds that intuition carefully and without heavy mathematics. You will generate embeddings for a few dozen sentences, plot them in two dimensions, and watch synonyms cluster while unrelated topics drift apart. You will see what happens to a negation, to a very long passage, to a proper noun the model has never seen, and to text in another language. Every claim is something you observe rather than something you accept.
Then you build search. Cosine similarity by hand first, then an index, then a proper comparison against keyword search on a dataset where each approach wins different queries. That comparison leads naturally into hybrid search, which is what most production systems actually use, and into reranking as a second, more expensive pass over the top results.
The final modules are practical: choosing an embedding model and dimension, handling updates when a document changes, keeping costs sane on a large corpus, and using embeddings for jobs other than search such as clustering support tickets and detecting near-duplicates. Storage is PostgreSQL with pgvector throughout.