Embeddings: How AI Maps Meaning Like a Bodega Shelf

An embedding is just a number — or a list of numbers — that stands in for a word, a sentence, or a whole document. You feed text into a model, and it spits out a vector, a long string of numbers. That vector is the embedding.

Think of it like the bodega shelf. Abuela doesn't need a spreadsheet to know where the sofrito lives. She walks in, smells the garlic, and her brain maps "sofrito" to a spot on the shelf. The shelf isn't the sofrito. It's a map. The embedding is that map — a compact representation of meaning.

Why does this matter? Because numbers can be compared. Two embeddings close together in math space mean two things are close in meaning. "Gato" and "perro" sit near each other. "Gato" and "taxi" sit far apart. The model doesn't know Spanish or English. It only knows proximity in a high-dimensional space. Thousands of dimensions, actually. You can't picture it. You just trust the math.

This is how search engines find your cousin's post without matching exact keywords. How chatbots understand that "la migra app" and "the immigration tracker" point to the same thing. How recommendation engines know you might like that recipe video after watching one about tamales.

Embeddings turned language into something a machine can measure. They're the bridge between human meaning and machine math. You don't need to understand the vector to use the result. You just need to know what the map gets you.

Next time your search returns something weird, ask: what words did I leave out that the embedding was trying to fill in?