What is Embedding?

The plain definition

An embedding model maps content — a sentence, a document, an image, a product — to a fixed-length vector, typically a few hundred to a few thousand numbers. The vector has no individually interpretable dimensions. What matters is relative position: content with similar meaning lands near other similar content.

That single property is unusually useful, because it converts a fuzzy human question ('what else is like this?') into a well-defined mathematical one (nearest neighbours by cosine similarity). Keyword search matches strings; embedding search matches meaning, so 'how do I cancel my subscription' retrieves a document titled 'ending your plan' despite sharing almost no words.

What they are actually used for

Semantic search and retrieval. The dominant use, and the mechanism underneath retrieval-augmented generation. Embed the corpus once, embed the query at request time, return nearest neighbours from a vector database.

Deduplication and entity resolution. Near-duplicate records defeat exact hashing entirely; embedding distance catches them. This matters for data quality generally and specifically for preventing duplicate leakage across training data splits.

Recommendations. Represent items and user histories in the same space, then recommend by proximity. This handles cold-start far better than collaborative filtering, because a brand-new item still has a content embedding.

Clustering and classification. Embeddings are strong general-purpose features. Freeze the embedding model, train a small classifier on top, and you have a competitive model from a few hundred labelled examples — the cheapest form of transfer learning.

The mistakes that cost the most

Mixing embedding models. Vectors from different models are not comparable — not even different versions of the same model. If you re-embed with a new model, you must re-embed the entire corpus. Teams discover this when retrieval quality collapses inexplicably after a routine upgrade.

Chunking badly. For retrieval, what you embed matters more than which model you use. A whole 40-page document averaged into one vector is too diffuse to match anything precisely; a single sentence lacks the context to be interpretable. Chunk on semantic boundaries — sections, headings — with modest overlap, and keep enough surrounding context that a retrieved chunk stands alone.

Assuming similarity means relevance. Cosine similarity measures topical resemblance, not usefulness. A question and a near-identical question are highly similar; the question and its answer may not be. This is why serious retrieval pipelines add a reranking stage, and why hybrid search — combining embeddings with keyword matching — reliably outperforms either alone, particularly on product codes, names, and exact identifiers where lexical matching is simply correct.