Embedding

Overview

An embedding turns something with meaning into a list of numbers a machine can compute with. The trick is that the numbers are not arbitrary: they are arranged so that items with similar meaning end up near one another in the resulting space. Two sentences about the same topic, or two images of the same kind of object, land close together, while unrelated ones sit far apart. This converts the slippery notion of "similarity" into simple geometric distance.

How it works

A neural network is trained to map inputs into this vector space such that meaningful relationships are preserved as spatial ones. Once trained, the model can embed any new input on demand, producing its coordinates in that space. Comparing two items then reduces to measuring the distance or angle between their vectors, a fast and cheap operation, which is what makes embeddings so useful as a foundation for search and retrieval.

Why it matters

Embeddings are the connective tissue of modern AI systems. They power semantic search, recommendation, clustering, and crucially the retrieval step in retrieval-augmented generation, where a query and a knowledge base are compared as vectors to surface relevant information. Generating and comparing embeddings at scale is a GPU workload in its own right, sitting upstream of many of the applications organizations actually deploy.