Embeddings
Numeric vectors that represent the meaning of text, images or other content, so that similar items end up close together and can be compared mathematically.
Definition
An embedding is a list of numbers, typically several hundred to a few thousand long, produced by a model that has learnt to place similar content near each other in that numeric space. Two sentences that mean roughly the same thing get vectors that point in roughly the same direction, even if they share no words. Similarity is measured with simple maths, most often cosine similarity.
Embeddings are produced by dedicated embedding models, not by the chat model itself. The same idea applies to images, audio and code, and multimodal embedding models can map text and images into one shared space.
Example: "refund not received" and "still waiting for my money back" contain no common keywords. Their embeddings are close, so a semantic search for one will find support tickets phrased like the other.
What it means in practice
- They power search by meaning, so customers and staff find things without guessing the right words.
- They are the first step of every RAG system; a poor embedding model limits everything downstream.
- They make clustering, deduplication and recommendations possible on unstructured data you already hold.
- Embedding is cheap per document but must be redone if you switch models, so the choice has long-term weight.
How Teksmobile uses it
Our embeddings strategy service benchmarks candidate models on your own documents and queries before anything is indexed. The results feed directly into vector search and knowledge-base builds, where we store vectors in pgvector or Milvus depending on scale.
Related terms
Part of the Teksmobile AI Glossary. For longer reads, see our guides.
Frequently asked questions
Do embeddings contain the original text?
Not directly. A vector cannot be read back as text, but it does encode meaning, and research has shown partial reconstruction is sometimes possible. Treat embeddings of sensitive content as sensitive data.
Can I change embedding model later?
Yes, but you must re-embed every document, because vectors from different models are not comparable. Plan the index so a full rebuild is routine. See choosing embedding models.
How many dimensions do I need?
Common models produce between a few hundred and a few thousand dimensions. More is not automatically better; retrieval quality on your own data, measured with evals, is the only reliable guide.