Embedding

An embedding model converts input into a fixed-length vector, often with hundreds or thousands of dimensions. Similarity between two items is then measured mathematically, for example with cosine similarity. Word embeddings such as word2vec (Mikolov et al., 2013) showed that this approach captures relationships between words; modern models embed whole sentences, documents and images.

Why it matters: embeddings power semantic search, recommendations, duplicate detection, classification and the retrieval step of RAG. If a tool promises to "search by meaning" or "find similar items", it almost certainly uses embeddings. Their quality, language coverage and cost affect how well that feature works, especially for non-English content.

Example: a support search built on embeddings can match the query "can't log in" to an article titled "Resetting your password", even though they share no keywords.

Limits: embeddings capture general similarity, not exact facts, and can miss rare terms, product codes or names. Many systems combine them with keyword search for better results.