What are embeddings?
Knowledge, data and integration
Embeddings are numerical representations of content such as text, images or audio, stored as lists of numbers called vectors. Items with similar meaning sit close together in this number space, so software can compare them by meaning rather than exact words. Embeddings power semantic search, clustering, recommendations and retrieval, and they are the foundation layer for knowledge work and retrieval-augmented generation. They are produced by a machine learning model, are not human-readable, and the same model must be used for both indexing and querying.
Reviewed by Jackie, Head of Learning & Development, Levellers · Last reviewed 8 June 2026
What this means
An embedding turns a piece of content into a list of numbers that captures its meaning and characteristics. Think of it as giving every document, sentence or image a precise location on a very large map. Items about similar topics end up near each other, even when they share no words in common. A query about maternity leave can sit close to a document about parental leave, because the model places them near each other based on meaning.
This matters because computers handle numbers far better than they handle language. Once content becomes numbers, software can measure how close two items are and rank them by similarity at speed and scale. The list of numbers is long, often several hundred or more values per item. That length is called dimensionality, and it gives the model room to record many shades of meaning. The numbers themselves are not meaningful to a human reader, and you cannot reconstruct the original text from them.
Why it matters
Embeddings turn unstructured content into something comparable. That reduces dependence on exact keywords, hand-built taxonomies and manual tagging, all of which are slow to maintain and brittle when language varies. For a business sitting on years of documents, tickets, proposals and policies, embeddings make that material searchable by meaning.
They are also the quiet engine behind most current knowledge tools. Semantic search, recommendation, deduplication and retrieval-augmented generation all rest on the same step: convert content to vectors, then compare. The prize is real. McKinsey's analysis notes that, going back to a 2012 estimate, knowledge workers spent about a fifth of the working week, roughly one day, searching for and gathering information, and embeddings are central to the tools that aim to cut that time.
How it works
From content to vectors
A trained model reads a piece of content and produces its vector. Early word-level methods such as word2vec showed that meaning could be captured as vectors and that related words sat close together. Modern sentence and document models, building on the Sentence-BERT line of work, produce a single vector for a whole passage so that similar passages can be compared directly using a measure such as cosine similarity.
Comparing by distance
Once items are vectors, similarity becomes a matter of distance. Closer vectors mean more similar meaning. A search system embeds the query with the same model used for the stored content, then returns the nearest items.
Chunking
Large documents are usually split into smaller passages, or chunks, before embedding. A single vector for a fifty-page report blurs too many topics together, so retrieval works better when each chunk covers one idea.
Storage and the same-model rule
Vectors are held in a vector database or a vector-capable search index. One rule is non-negotiable: use the same embedding model to index your content and to embed queries. Vectors from different models are not comparable, because each model builds its own number space.
Examples
Knowledge search: an employee asks a question in plain language and the system returns the most relevant passages from internal documents, even where the wording differs.
Ticket clustering: incoming support tickets are embedded and grouped by similarity, surfacing recurring themes without anyone tagging them by hand.
Proposal reuse: a sales team finds previous proposals close in meaning to a new brief, so strong sections can be reused.
Recommendation and matching: products, candidates or articles are matched to a profile or query by closeness in vector space.
Common misunderstandings
"An embedding is a summary." It is not. It is a numerical fingerprint of meaning, not a shortened version you can read.
"Embeddings understand truth." They capture statistical patterns of meaning, not facts. Two false statements that sound alike will sit close together.
"Any model works with any other." Vectors are only comparable when produced by the same model. Mixing models breaks retrieval silently.
"Embeddings remove the need for metadata." They do not. Dates, permissions, source and document type still matter for filtering and trust.
Risks and boundaries
Fit for purpose: a general model may not capture specialist language well. Test on your own content before committing.
Chunking choices shape quality. Chunks that are too large or too small both hurt retrieval.
Model change is costly. If you switch embedding models, you must re-embed everything, because old and new vectors are not comparable.
Embeddings do not solve permissions, provenance or freshness. They tell you what is similar, not who is allowed to see it, where it came from, or whether it is current. Those controls sit elsewhere in the system.
What to do next
Pick a representative set of your own content, perhaps a few hundred documents covering the topics you care about. Generate embeddings with one model. Inspect the nearest items for a handful of real queries and judge whether the matches make sense to a person who knows the material. That small test tells you more about fit than any vendor benchmark. Keep metadata alongside the vectors from the start, and decide early which model you will standardise on, since changing it later means re-embedding everything.
Have a question or a suggestion, or want to understand how we research and review these guides? Read about our editorial standards and how to reach us.
FAQs
Are embeddings only for text?
No. The same idea applies to images, audio and other content. Each needs a model trained for that type, and you should not mix vectors across types unless the model is designed for it.
If two items have similar embeddings, are they the same?
No. Similar means close in meaning, not identical. Closeness is a matter of degree, and you choose how close counts as a match.
Can I compare embeddings from two different models?
No. Each model builds its own number space, so vectors from different models are not comparable. Use one model for indexing and querying.
Where are embeddings stored?
In a vector database or a vector-capable search index, usually with metadata such as source, date and access rules held alongside.
Do embeddings replace a taxonomy?
They reduce reliance on manual tagging, but structured metadata and categories still help with filtering, governance and trust.
How long is an embedding?
It is a fixed list of numbers, often several hundred or more values. That length is called dimensionality and is set by the model.
Do I need to re-embed when I add new documents?
New content must be embedded with the same model you already use. You only re-embed everything if you change the model.
Are embeddings human-readable?
No. They are numbers, and you cannot read the original content back from them.
Sources
Efficient Estimation of Word Representations in Vector Space (arXiv (Mikolov, Chen, Corrado, Dean)). Foundational word2vec paper establishing words as vectors with meaningful distances.
Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (ACL Anthology (Reimers and Gurevych)). Foundational sentence-embedding method enabling fast similarity comparison of passages.
ISO/IEC 22989:2022 Information technology, Artificial intelligence, AI concepts and terminology (ISO/IEC). Standardised definitions for AI concepts and terminology used in the explainer.
The economic potential of generative AI: The next productivity frontier (McKinsey & Company). Evidence that knowledge workers spend about a fifth of the week searching for information, and the value of better knowledge tools.