Request a call

Embeddings

Definition

Embeddings are lists of numbers that capture the meaning of text or images, so software can measure how closely related one item is to another.

Updated

Also called vector embeddings.

How it works

An embedding model reads a passage and outputs a vector: a fixed-length list of numbers, usually between a few hundred and a few thousand of them. Passages with similar meaning get vectors that point in similar directions. "Cancel my subscription" and "how do I stop my plan renewing" share almost no words, yet they land close together. Closeness is usually measured with cosine similarity, the measure OpenAI's embeddings guide recommends.

Embedding models are separate from chat models and far cheaper to run. OpenAI, Google, Cohere and Voyage AI sell them through APIs, and open models can run on your own hardware.

How businesses use it

Embeddings are the plumbing under several things buyers ask for by other names:

  • Semantic search over a help center or intranet
  • Retrieval in RAG assistants
  • Grouping thousands of support tickets or survey comments by theme
  • Matching records that describe the same thing differently, such as supplier names or product listings
  • Routing, by comparing an inbound email with labeled examples of each category

Because embedding is cheap, running your whole document library through a model is rarely the expensive part of a RAG build. Cleaning and parsing the documents usually is.

Where it goes wrong

Vectors from different models are not compatible. Change the embedding model and you must re-embed everything, so record which model and version produced each vector.

General-purpose models can miss the vocabulary of a specialist field, such as part numbers or drug names. Test retrieval on your own questions before committing to a model, and add keyword search where exact terms matter (hybrid search).

Embeddings also capture meaning loosely. "Refund approved" and "refund denied" can sit close together, because they are about the same topic. Retrieval finds candidates. A model or a rule still has to read them and decide.

Related terms

  • Vector database

    A vector database stores embeddings, the numeric fingerprints of text or images, and quickly finds the stored items whose meaning is closest to a query.

  • Semantic search

    Semantic search finds results by meaning instead of exact words, so a query can match relevant content even when they share no vocabulary.

  • Chunking

    Chunking is splitting documents into smaller passages before they are embedded and indexed, so a retrieval system can find and return the exact part that answers a question.

  • Retrieval-augmented generation

    Retrieval-augmented generation is a method where a system first searches your own documents for relevant passages, then gives them to a language model to answer from.

All 30 definitions are in the AI and search glossary.

Find out where this fits in your own systems

Bring one process or page to a 30-minute call.

Request a 30-minute scoping call

We reply within 1 working day and suggest times that suit you.