Meaning as coordinates
An embedding converts the meaning of a word, a passage, or an image into a sequence of hundreds or thousands of numbers — a vector. Words close in meaning, such as king and monarch, are designed to land close together. Thinking of it as turning meaning into coordinates on a map is a useful picture.
Why it matters: search by meaning
Once text has coordinates, closeness in meaning becomes distance you can compute. That is the decisive difference from keyword search.
“How to claim expenses” and “getting back money you paid out of pocket” share almost no words but land close together as embeddings. That is how material matching your intent can be found despite different wording — semantic search.
Where it is used
- RAG: finding the passages closest in meaning to a question and passing them to the model
- Internal document search and automated FAQ responses
- Similar-product recommendations on e-commerce sites
- Classifying text and grouping similar enquiries
The dedicated store for holding large numbers of embeddings and searching them quickly is the vector database. Reading this alongside the RAG entry connects the whole picture of an internal knowledge assistant.
How it relates to LLMs
LLMs themselves convert text into embeddings before processing it. Embeddings are therefore both a tool for search and the underlying mechanism by which a model handles language at all. Knowing that AI works with language as numeric coordinates makes technical writing about AI considerably clearer.
Conversion uses dedicated embedding models, which vary in language coverage. When RAG retrieval is performing poorly, the choice of embedding model is worth revisiting alongside how documents were chunked.
Turning language into coordinates is the starting point for AI search, recommendation, and classification alike — infrastructure for the generative AI era, even if it stays out of sight.