Embeddings are numerical representations that capture the meaning of a piece of data (a word, a phrase, a document, or an image) in the form of a vector with hundreds or thousands of dimensions. The key is that these vectors are constructed so that elements with similar meanings remain close in the vector space, while very different ones remain far apart. Thus, the model does not work with raw text, but with coordinates that encode semantic relationships.
Their importance lies in the fact that they allow computers to compare meanings mathematically, measuring, for example, the distance or the cosine between two vectors. This enables tasks such as:
- Semantic search: finding results by meaning, not by exact words.
- Recommendation systems and grouping (clustering) of similar content.
- RAG, where relevant fragments are retrieved to feed a language model.
A practical nuance: the same term can have different embeddings depending on the model that generates them, so it is advisable to always use the same model for indexing and querying, ensuring that the vectors are comparable to each other.