Embeddings and search concept

What are embeddings?

Reference

What are embeddings? They are numerical representations of data, such as text, arranged as lists of numbers called vectors. A model produces a vector so that, for a chosen task, items that are alike in useful ways can be placed near one another in the vector space. Comparing those vectors can help find related text, but closeness is a model’s task-specific signal, not a guarantee that two texts mean the same thing.

Reviewed 2026-10-09

An embedding is a learned numerical representation

An embedding is a vector: an ordered list of numbers that represents an item. The item might be a word, a sentence, a document, or something other than text. A model maps the input into a space with many dimensions; each number is a coordinate in that space. The coordinates are useful because a computer can compare them mathematically, even when the original items are difficult to compare directly.

The dimensions usually do not have simple labels like “topic” or “tone.” Their meaning depends on how the model was trained and what task the representation is meant to support. So an embedding is not a neat dictionary entry or a full, human-readable account of a text. It is a compact representation that can help a particular system find patterns.

How text becomes an embedding

To create a text embedding, an embedding model processes text and returns a vector of numbers. The text might be a query or a passage in a document collection. The model’s training shapes which texts end up near one another: systems built for different tasks can represent the same text differently. A model may produce an embedding for a whole sentence or for another chosen piece of text.

A search system can embed its passages ahead of time, then embed a query when someone searches. It compares the query vector with passage vectors using a similarity measure, such as cosine similarity, and retrieves close matches. This depends on the query and passages being represented in a compatible way. The resulting score measures similarity in that representation; it does not, by itself, establish that a retrieved passage is correct or answers the query.

Where embeddings fit in search

An embedding is the representation; vector search is a way to find vectors that are close to a query vector. A vector index can help search those representations, and a vector database can store and retrieve them. In semantic search, this can help find passages that relate to a query even when they use different wording. These neighbours are part of representations and retrieval methods.

Keyword search looks for query terms in text, which can be useful when exact wording matters. Embedding search can help with related wording, but its similarity scores can miss useful results or rank irrelevant ones highly. Hybrid search combines keyword and semantic results. Search systems may also rerank an initial set of candidates, but embeddings remain one part of the retrieval process, not a measure of whether a claim is true.

What embeddings do not tell you

A common confusion is to treat an embedding as the meaning of a sentence, stored in a form a person could read. It is better understood as numbers shaped to support a task. Nearby vectors indicate similarity under a chosen model and comparison method; they do not prove that the texts are interchangeable, factually consistent, or equally useful. The individual coordinates may also be hard to interpret.

Another confusion is that all embedding models produce the same vector for a given text. The representation depends on the model and its training, so vectors from different models are not automatically comparable. Even within a compatible set, a similarity score is not a truth score. For search, check whether the retrieved text is relevant to the query and whether it actually supports the answer being made.

Questions people ask

How do text embeddings work?

An embedding model processes text and outputs a vector of numbers. A search system can compare a query’s vector with vectors for stored passages using a similarity measure, then retrieve nearby matches. The model and its training shape what counts as close.

Are embeddings the same as meaning?

No. They are numerical representations shaped for a task. Similarity between vectors is evidence that the model places two items near each other, not proof that their meanings are identical or that either text is true.

How are embeddings different from keyword search?

Keyword search matches terms in text. Embedding search compares vector representations and can retrieve related wording. Each can miss useful results, so some search systems combine both approaches.

Learn more

Embeddings: Embedding space and static embeddings 45 min
An introductory explanation of vectors, embedding spaces, and why distance between embeddings can represent relative similarity.
Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks
A freely available research paper showing how sentence embeddings can be compared for semantic similarity and search.

What changed

Sources

Part of Embeddings and search: Representations and retrieval methods