Architecture
Embedding
Text, images or products are translated into long series of numbers so that similar things lie close together: “invoice” and “payment” end up as neighbors even though not a single letter matches. This map of meanings underlies semantic search, duplicate detection, recommendation — and almost every RAG system. Embeddings are unspectacular and carry a surprising amount of infrastructure.
An example
An embedding turns a sentence into a long series of numbers that represents its meaning. The practical effect: “How do I cancel my contract?” and “contract termination notice period” end up close together even though they share not a single word. A full-text search does not find one from the other; a search over embeddings does. That is exactly why search in manuals and knowledge bases suddenly works where it did not for years.
A closer look
An embedding is a vector with a few hundred to a few thousand dimensions in which closeness means similarity of content. Search thereby becomes a calculation of distance: “notice period” and “end of contract” lie close together even though they share no word.
Why it matters
Embeddings are the most unspectacular and at the same time the most reliable building block of modern AI applications — they do not hallucinate, because they generate nothing. Two things are regularly overlooked all the same. First: an embedding is not encryption. The original text can be reconstructed from the numbers to a considerable extent; a vector database containing personnel data is, under data protection law, a database containing personnel data. Second: if you change the embedding model, the entire stock has to be recomputed. Old and new vectors are not comparable.
Commonly misunderstood
Similarity is not correctness. An embedding finds the text that matches best, even if that text is outdated, refuted or simply wrong.