The problem: a computer works with text as a set of characters, without understanding meaning. Ordinary search looks for word matches — and misses when the meaning is the same but the words differ. Example: a user searches "how to return a purchase", and in your base an article is called "processing a product return". By words they do not match ("return a purchase" is not "product return" is not "processing"), and ordinary search will not find the needed one, though it is exactly what the person searched for. Embeddings solve this. The idea: turn text into a set of numbers (a vector) reflecting MEANING. The key property: texts with close meaning get close number-vectors, even if the words differ. "Return a purchase" and "product return" will have close embeddings because the meaning is close. Now you can search by MEANING: turn the query into an embedding, search for texts with the nearest embeddings — find the relevant regardless of exact words. Where it is used (embeddings are the foundation of much in AI): 1) Semantic search — find by meaning, not words. 2) RAG — that very indexing and search of fragments by meaning (from the RAG course). 3) Recommendations — find similar products/articles/content. 4) Clustering and analysis — group the similar, find connections. 5) Duplicate/similar detection. Analogy: an embedding is like the "meaning coordinates" of a text on an imaginary map. Texts about similar things are near each other, about different things far apart. This gives the computer a way to "understand" meaning proximity through number proximity. Practically: embeddings turn the problem of "understand meaning" into the problem of "compare numbers", which a computer solves easily and fast. This is one of the fundamental techniques underlying semantic search, RAG, recommendations. This course covers embeddings and vector databases (where they are stored and searched) in detail.