Embeddings

How it works

An embedding is a numeric vector that represents the meaning of a piece of text (or an image, or audio) as a point in high-dimensional space, so that similar meanings sit close together. A model converts a sentence into a list of hundreds or thousands of floating-point numbers; the geometric distance between two embeddings approximates their semantic similarity. This is what lets software search by meaning rather than by exact keywords: you embed the query, embed the documents, and compare vectors.

When it matters

Embeddings matter wherever you need semantic search, clustering, recommendation, or retrieval. They are the first step of a RAG pipeline: you embed both the query and your documents, then look up the closest matches in a vector database. Get the embedding model and chunking strategy right and retrieval quality follows; get them wrong and the model grounds its answer in the wrong passages.

Common misconceptions

• "Embeddings store the original text." They store a representation of meaning, not the text; you keep the source separately.
• "Any embedding model works for any task." Quality varies by domain and language; evaluate on your own data.
• "More dimensions is always better." Larger vectors cost more storage and compute for marginal gains past a point.