Representation learning is the branch of machine learning concerned with training models to automatically discover useful numerical representations of raw data, rather than relying on features hand-crafted by humans. The embeddings that power vector search are a direct product of representation learning.
The core idea is that a well-chosen representation makes downstream tasks far easier. Instead of an engineer manually deciding which features of an image or document matter, a representation-learning model learns, from large amounts of data, to map inputs into a space where meaningful structure emerges — where similar things are close and the dimensions capture salient properties. This learned space is exactly the latent space that embeddings inhabit.
Representation learning is what makes modern semantic search possible. Embedding models like BERT, sentence transformers, and CLIP are all representation learners, trained so that their output vectors encode meaning in a form amenable to similarity comparison. Understanding vector databases as infrastructure for storing and searching learned representations connects them to this broader and deeply influential area of machine learning.