What are embeddings?

The idea in plain language

Picture a map where every sentence you have ever written sits somewhere. Sentences about cooking cluster in one area, sentences about taxes in another. Now make that map have hundreds of dimensions instead of two. An embedding is a sentence's coordinates on that map, usually a list of a few hundred to a couple of thousand numbers.

A trained model produces those coordinates. Nobody hand-assigns what each number means; the model learns positions that put related content close together. To compare two things, software measures the distance or angle between their embeddings, most often with a score called cosine similarity.

Text, images and audio

Embeddings are not limited to text. Image models embed pictures, audio models embed sounds and voices, and some models are trained to put text and images into the same space. That shared space is what lets you type “dog on a beach” and find a photo that has no caption at all: the words and the picture land near each other.

Here is a small example. Three captions, “a cat sleeping on a sofa,” “a kitten napping on the couch” and “a spreadsheet of Q3 revenue,” are embedded. The first two land very close together despite sharing only one word; the third lands far away. A search for “pet asleep on furniture” lands next to the first two and returns them, while the spreadsheet never appears.

What they are used for

• Semantic search: embed the query, return the nearest stored items. See semantic search.
• Grouping and recommendations: “more like this” is just the nearest neighbors of an item.
• Duplicate detection: two near-identical photos have near-identical embeddings. See near-duplicate detection.
• Voice matching: voice embeddings help tell speakers apart in speaker diarization.
• Grounding AI answers: retrieval-augmented generation uses embeddings to pick which passages a language model reads.

Limits

Embeddings capture general meaning, so they can blur distinctions that matter to you, such as “increase” versus “decrease,” or two products with similar descriptions. They are also only comparable within one model: embeddings from two different models live on different maps and can't be mixed. Switching models means re-embedding everything. And a long document squeezed into one embedding loses detail, which is why search systems usually embed short chunks.

In MediaFind

MediaFind computes embeddings on your computer for transcript passages, video keyframes and photos, and stores them in your local library. They power search by meaning, visual search, related clips and reverse image search. No text or image is sent anywhere to be embedded.

Frequently asked questions

Are embeddings the same as vectors?

An embedding is a vector, meaning a list of numbers. The word embedding emphasizes that the numbers were learned to capture meaning.

Can I compare embeddings from different models?

No. Each model builds its own space, so similarity scores only make sense between embeddings from the same model.

Do embeddings need a GPU?

Not necessarily. Smaller embedding models run fine on a laptop CPU, and a GPU mainly makes large batches faster.