Embeddings
Numerical representations of text that capture meaning.
Explanation
Embeddings convert text into number vectors that capture semantic meaning. Similar concepts are close to each other in vector space, regardless of the exact formulation.
How it works
A specialized model reads a text and outputs a vector (e.g. 1536 numbers). "Dog" and "puppy" have similar vectors, "dog" and "car" do not.
Example
In a RAG pipeline, all documents are saved as embeddings. When a user query is made, the embedding of the question is compared with the document embeddings.
Why it matters
Embeddings enable semantic search - the system finds relevant information even if other words are used.
What are embeddings in AI?
An embedding represents a piece of content, such as a sentence, a document or an image, as a list of numbers called a vector. Each number captures some property the model learned during training. Individually the values are not interpretable, but together they encode meaning.
The practical benefit is that similarity becomes measurable. The distance between two vectors, usually computed as cosine similarity, shows how close two pieces of content are in meaning. “Notice period for my lease” and “How early do I have to cancel my apartment?” end up close together even though they share almost no words.
How businesses use embeddings
The most common use is semantic search in RAG systems: documents are split into passages, stored as embeddings in a vector database and retrieved by meaning rather than keywords. Other uses include clustering similar tickets or articles, duplicate detection, recommendations and text classification.
When choosing an embedding model, three things matter most: how well it handles your language and domain vocabulary, where it runs (cloud API or locally) and the vector size. Larger vectors are often slightly more accurate but need more storage. Important: all documents and all queries must be embedded with the same model. Switching models means re-embedding the entire corpus.
Frequently asked questions
What is the difference between embeddings and an LLM?
An LLM generates text. An embedding model does not generate text but produces a vector that represents the meaning of an input. In RAG systems, embeddings find the right documents and the LLM writes the answer.
How many dimensions does an embedding have?
It depends on the model. Common sizes range from a few hundred to a few thousand dimensions, for example 768, 1,024, 1,536 or 3,072.
Are embeddings only used for text?
No. There are embedding models for images, audio and code, as well as multimodal models that place text and images in the same vector space.