Embeddings
Quick definition
An embedding is a list of numbers (a vector) that represents the meaning of a text, image or other item, so that similar items get nearby vectors. Embeddings are what makes search by meaning possible.
On this page
How embeddings work
An embedding model reads a text and outputs a fixed-length vector, typically several hundred to a few thousand numbers. The model is trained so that texts with related meanings land near each other in that space: “refund policy” and “how do I get my money back?” share almost no words but produce close vectors.
The idea goes back to word vectors such as word2vec (Mikolov and colleagues, 2013), which learned that words appearing in similar contexts have similar meanings. Sentence-level models such as Sentence-BERT (2019) extended it to whole sentences and passages, which is what search and question answering need.
What they are used for
Embeddings power semantic search: compare the vector of a question with the vectors of stored passages and return the closest. They are the retrieval step of retrieval-augmented generation (RAG). They also serve for clustering similar conversations, detecting duplicates, recommending similar products and routing a message to the right topic.
Multilingual embedding models place a question in French and a passage in English close together when they mean the same thing, which lets one knowledge base serve customers in many languages.
Practical points
Vectors from different models are not comparable: if you change the embedding model, every stored passage has to be re-embedded. Embeddings capture meaning, not exact values, so they are poor at precise filters such as a price range or a product reference — those need structured fields or keyword search alongside. And long documents must be split into passages before embedding, because one vector for a whole manual blurs its content.
Embedding models also differ in language coverage, dimension (and therefore storage cost) and where they are hosted.
Embeddings in AI SmartTalk
Each AI SmartTalk assistant uses one embedding model for all of its knowledge. New assistants use an EU-hosted multilingual model by default, and OpenAI embedding models are also available. The vectors are stored in PostgreSQL with pgvector and combined with keyword and trigram search, so that product references and exact terms are found as reliably as paraphrased questions.
Frequently asked questions
What is the difference between an embedding and a token?
A token is a piece of text, roughly a word fragment, that a model reads. An embedding is a vector of numbers that represents meaning — of a token, a sentence or a whole passage.
Are embeddings the same as a vector database?
No. Embeddings are the vectors; a vector database or index stores them and finds the closest ones quickly.
Can I switch embedding models later?
Yes, but every stored item must be re-embedded with the new model, because vectors from different models live in different spaces and cannot be compared.