Embeddings

An embedding is a fixed-dimension vector produced by a dedicated model (an embedder) from text, an image or another object. Proximity between two embeddings under a chosen metric (cosine, Euclidean, dot product) is read as semantic proximity of the underlying objects: the same idea expressed in different words yields nearby vectors.

Where it works: semantic search in a vector database, retrieving relevant fragments in a RAG system, recommendations, deduplication and entity linking, clustering of inbound, anti-fraud. In each of these, an embedding is not “magic” — it’s a way to reduce “similar in meaning” to arithmetic.

What gets overlooked: the embedder caps the quality of the entire retrieval. Swapping the model changes the geometry of the space: an index built on one set of embeddings does not work on another without a full rebuild. Vectors must be normalized for the chosen metric — otherwise results are random. Dimensionality and inference cost directly drive latency and budget, so the embedder is picked not as “the strongest” but as the strongest that fits the system’s constraints.