Embedding
A vector of numbers that represents a piece of text (or image, or audio) such that semantically similar items produce numerically similar vectors.
An embedding is a list of numbers — typically 768 to 3072 of them for text — produced by running content through an embedding model. The specific numbers aren't meaningful individually. What matters is the geometry: content with similar meaning ends up at nearby points in this high-dimensional space, and content with different meaning ends up far apart.
The reason this matters: once text is represented as coordinates, you can search by meaning instead of by keyword. "articles about people leaving their jobs" will match a document about "engineers who quit" without any word overlap. This is the mathematical foundation of modern semantic search and every RAG system.
Embeddings are cheap to compute (much cheaper than running a full LLM), so they're the workhorse of large-scale document indexing. A typical setup: compute an embedding for every chunk of a document store, save them in a vector database, then compute an embedding for each user query and find the nearest matches. The nearest matches become the context for the LLM to answer from.