.svg)
Kronehoefstraat 85
5612 HL Eindhoven
BTW NL863902509B01
What is it?
A vector embedding is a conversion of text or other data into a sequence of numbers: a vector in a mathematical space. The numbers are not arbitrary; they are calculated by a language model that has learned which words, sentences, and concepts are semantically similar to each other. Texts with similar meanings are positioned close together in that vector space; texts with very different meanings are positioned far apart.
Embeddings are the technical layer that makes semantic search, RAG, and recommendation systems possible. Without embeddings, a computer treats text as a sequence of characters; with embeddings, it treats text as a point in a space of meaning.
Why it matters for SMEs
Embeddings are the reason AI systems understand text rather than just matching it. For SMEs, that matters as soon as you want to use AI for searching, comparing, or retrieving information based on content.
- Embeddings allow an AI agent to compare documents, candidate profiles, or emails based on how similar their content is, without requiring you to use the exact same words as the source.
- They form the basis of RAG systems: by storing business knowledge as embeddings in a vector database, an AI model can give current and specific answers rather than generic ones.
- Quality control and classification of unstructured input, such as complaints, quote requests, or support tickets, become automatable because embeddings recognise semantic categories.
As an SME owner you will not typically work with embeddings directly, but every AI tool that searches, compares, or retrieves based on meaning uses them under the hood.
How it works
An embedding is created by passing a piece of text through an embedding model: a neural network trained to capture meaning relationships. The output is a vector, typically consisting of hundreds to thousands of numbers.
- A piece of text (word, sentence, paragraph, or document) is passed to the embedding model.
- The model converts the text into a fixed sequence of numbers: the embedding vector.
- This vector is stored in a vector database or memory.
- When a comparison or query is made, the new text is converted in the same way.
- The distance between the two vectors is calculated, typically as cosine similarity: the smaller the angle, the more closely the meanings align.
Commonly used embedding models include OpenAI embeddings (text-embedding-3-small, text-embedding-3-large), Cohere Embed, and open-source models from Hugging Face. The choice of embedding model partly determines how well semantic similarity is recognised for your specific language and domain.
Example in practice
Picture a housing association that wants tenants to be able to ask questions about their tenancy agreement through a chat interface. The administrator loads all tenancy agreements into the system: each contract is split into clauses and each clause is converted into an embedding. When a tenant asks "what is the notice period?" that question is also converted into an embedding. The system finds the clauses whose embeddings are closest to the query's embedding, locates the relevant notice provisions, and passes them to a language model that formulates a clear answer.
Comparison and misconceptions
An embedding is the vector representation of a specific piece of text; a vector database is the system that stores embeddings and makes them searchable. The difference is between a single data point and the library that manages all data points. Both are needed for a working semantic search system.
.svg){kind=icon}