**Embeddings** are dense numeric vectors that represent text in a way that captures meaning. The defining property: two pieces of text with similar meaning land at nearby points in the vector space, and unrelated pieces land far apart.
This matters because once meaning becomes a position in space, comparing meanings becomes geometry. To find documents related to a query, you embed the query, embed every document, and look for the documents whose vectors are closest to the query vector. To cluster sentences by topic, you run a clustering algorithm on the vectors. To classify a new piece of text, you compare its vector to vectors of labelled examples.
The other three options in the question describe different operations that have nothing to do with this. Compression makes text smaller, encryption hides it from unauthorised readers, translation rewrites it in another language. None of those produce a vector space where geometric proximity equals semantic similarity.
The vectors themselves are produced by an embedding model trained on huge amounts of text. The model has learned how to arrange new text on the map so that the arrangement respects meaning. That is what makes the geometry useful: when the model places vectors well, simple math on top recovers retrieval, clustering, and classification results that match human intuition.
**Embeddings** are dense, fixed-dimensional vectors that map discrete text into a continuous space where semantic similarity becomes geometric proximity.
A modern text embedding has a few hundred to a few thousand dimensions. The embedding model is trained so that semantically related texts end up close in that space, while unrelated texts end up far apart. Once that property holds, retrieval, clustering, classification, and recommendation all reduce to elementary operations: nearest-neighbour search, distance-based clustering, vector arithmetic.
Option B is correct. Option A confuses embeddings with compression, which preserves the original text byte for byte. Option C confuses them with encryption, which has the strong property that only key-holders can recover the input. Option D confuses them with machine translation, which produces target-language text rather than a vector.
The practical use case that has made embeddings ubiquitous in modern NLP is retrieval-augmented generation, where embeddings drive the retrieval step that fetches relevant context for an LLM to read. Without embeddings, retrieval over large corpora would have to rely on lexical keyword matching, which misses semantic matches and forces the system to depend on exact wording.
Quick reference
TL;DR
Embeddings map discrete text into a continuous vector space where geometric distance corresponds to semantic similarity, turning meaning into math.
Imagine giving every sentence a spot on a giant map, where sentences that mean similar things end up near each other. "My dog is barking" and "the puppy is loud" would sit close by. "The cake is tasty" would sit far away on a different part of the map. Once every sentence has a spot, comparing meanings becomes as easy as measuring how far apart two pins are. That is what this idea does for text. It turns words into numbers a computer can line up on a map. You can sort, search, group, and pair up sentences by simple map distance. The trick is the program that picks the spots. It has read tons of text and learned how to place new sentences in a place that respects meaning, so the math we do on top actually matches what humans think feels close.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Exact lexical matching, like finding documents containing a specific phrase. Embeddings approximate meaning, not exact wording. A BM25 keyword index is the right tool for verbatim queries.
Tasks requiring exact byte-level reconstruction of the input. Embeddings are lossy by design. They preserve meaning but not the original text. Use compression if reconstruction is the goal.
Privacy critical pipelines where original text must remain hidden. Embeddings can leak surprising amounts of information about the input, including exact phrases in some cases. Use proper encryption, not embedding as pseudo encryption.
Quick scan · 2–4 min
Everything important, quickly.
Define embeddings as dense vectors that encode meaning geometrically. Identify nearest-neighbour search as the operation that makes the structure useful. Dismantle each distractor with one sentence linking it to a different concept (compression, encryption, translation). Close with RAG and clustering as the main applications that make embeddings central to modern NLP.
Real-world examples
Real products, models, and research that use this idea.
1OpenAI's `text-embedding-3-large` produces 3072-dimensional vectors and is the default embedding for many RAG stacks in 2026.
2Voyage AI's `voyage-3-large` provides domain aware variants for code, finance, and law where general embeddings underperform.
3Pinecone, Weaviate, Qdrant, and pgvector all store millions of embeddings and serve nearest-neighbour queries in milliseconds via approximate ANN indexes.
4Anthropic's contextual retrieval system uses embeddings as the first stage before a reranker, demonstrating the typical two stage retrieval pattern.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhat is the difference between word embeddings (Word2Vec, GloVe) and modern sentence embeddings?
Sign in or go Pro to see the approach for this follow-up.
QHow does retrieval-augmented generation use embeddings, and what other components does the pipeline need?
Sign in or go Pro to see the approach for this follow-up.
QWhat is the relationship between embedding dimensionality and quality, and why have Matryoshka embeddings become popular?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Confusing embeddings with compression, encryption, or translation. Compression is lossless, encryption is reversible by key, translation produces text. Embeddings produce dense vectors organised by meaning, not by any of those properties.