---
title: Embedding
description: 'A numeric vector produced by a representation model so useful relationships among items can be compared with geometric operations.'
canonical_url: 'https://darkfactory.dev/glossary/embedding'
markdown_url: 'https://darkfactory.dev/glossary/embedding.md'
collection: glossary
date_published: '2026-08-03T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Embedding


## Definition

An embedding is a numeric vector produced by a representation model so useful relationships among items can be compared with geometric operations. Text passages, images, audio clips, products, users, graph nodes, and code can all be embedded when a model has been trained or configured to represent them.

The geometry reflects the model's training objective and data. Vectors that are close under cosine similarity may be similar for the task the model learned, but they are not necessarily interchangeable, factually consistent, or equally useful for another task. Distance also depends on normalization and the chosen metric.

## Development of modern embeddings

Distributed vector representations predate current embedding services. In 2013, Tomas Mikolov and coauthors introduced efficient word2vec architectures for learning continuous word vectors from large text collections and evaluated the resulting syntactic and semantic relationships. Later systems extended the approach to sentences, documents, images, audio, and shared spaces across modalities.

This is a technical lineage, not a claim that word2vec coined the word "embedding" or invented every distributed representation.

## Distinguish it from nearby terms

- An **embedding model** produces the vector. The embedding is one output from that model.
- A **vector database** stores vectors, builds indexes, and performs searches. It does not determine what the geometry means.
- **Semantic search** retrieves items using an embedding and similarity method. It is an application, not the representation itself.
- A **feature vector** is any ordered set of model inputs. An embedding is usually learned so relationships can be expressed in its vector space.

## Operational significance

Record the embedding model and version, vector dimension, preprocessing, distance metric, normalization, and index settings. Changing any of them can reorder retrieval results. Thresholds should be calibrated on the actual task because a cosine score has no universal meaning across models or datasets.

## Check your understanding

Two policy documents have nearby embeddings because they discuss the same topic, but one supersedes the other. The similarity helps retrieval; it does not tell the system which policy is current.

## Also called

vector embedding

## Related factory areas

- [Context, memory, knowledge & skills](https://darkfactory.dev/factory/context-memory-skills)

## Evidence and further reading

- [Google Machine Learning Glossary](https://developers.google.com/machine-learning/glossary/)
- [Efficient Estimation of Word Representations in Vector Space](https://arxiv.org/abs/1301.3781)