Embeddings: Meaning as Numbers

An embedding model maps text, images, or other inputs to vectors — long lists of numbers. Items used in similar contexts tend to land near each other on the map. Real embeddings have hundreds or thousands of dimensions; humans can only look at two or three, so visualizations squash the real space down and show the rough neighborhoods.

Search systems compare these vectors to find candidates that may be relevant even when the words differ. Ask for “cheap evening meal” and an embedding search can surface a page about “affordable dinner” that shares zero keywords with your query. That meaning-level matching is what powers modern semantic search — and the retrieval half of RAG.

Simplified educational illustration — hand-placed 2D points, not real model output. Real embeddings have hundreds of dimensions and are learned, not assigned by hand.

The dashed lines tell the story: the query “quick lunch near the station” sits between the food cluster and the travel cluster, and its nearest neighbors mix both worlds — sandwich, bicycle, train. Nearness pulled in meaning from both sides of the question, even though the query shares no exact words with any point.

Our map uses hand-assigned food and transport scores — a toy version of the real thing. Try this thought experiment: where would you place “airport sandwich”? High on both axes, right in the middle. Where would “tire” go? Far toward travel, zero toward food.

Now imagine doing that for every word in a library, in a thousand dimensions, with positions learned from billions of sentences instead of assigned by hand. That is a real embedding model — and the reason “cheap evening meal” can find “affordable dinner” without sharing a single word.

Nearness is a clue, not proof of equal meaning or truth. Two sentences can sit side by side because they share a topic while disagreeing completely (“vaccines work” vs “vaccines don't work” can embed closely — they are about the same thing). Embeddings capture aboutness, not correctness.

The model, the distance measure, and the task all change which results are useful. Embeddings also inherit the biases of their training data — whatever associations the text contained, the map quietly preserves. Use them for candidate retrieval, then verify with the actual text.