Why vectors and embeddings
Similarity in number-space — the intuition behind search and retrieval.
Computers don’t “feel” that dog and puppy are related. Embeddings turn words, sentences, or documents into lists of numbers (vectors) so that similar meaning tends to sit nearby in that number-space.
Near = more similar meaning. Far = less related. No math required.
The intuition
Imagine placing sticky notes on a wall: related ideas closer together, unrelated ones farther apart. An embedding model learns a high-dimensional version of that wall. “Distance” (or angle) between vectors becomes a rough measure of similarity.
You don’t need the math. You need the idea: meaning → coordinates → compare by nearness.
Why we bother
- Search / retrieval — Find documents like your question without exact keyword matches (“refund policy” can find “money-back terms”).
- Recommendations — “More like this.”
- RAG-style apps — Retrieve relevant notes first, then ask a language model to answer using them (reduces pure guesswork).
See also: Glossary — Embedding.
Limits (keep it honest)
Nearby ≠ true. Two vectors can be close for the wrong reason. Embeddings don’t replace reading the source — they help you find candidates faster.
Takeaway
Vectors/embeddings are how modern AI systems measure likeness so they can retrieve and organize text (and other media) at scale. Inside a transformer, embeddings are also the starting “coordinates” for each token before attention mixes context.
Write three short phrases: two that mean nearly the same thing, and one that doesn’t. Mentally place them on a wall — the first pair close, the odd one out farther. That’s the embedding story without code.