What is embedding similarity?

It is a score of how close two texts are in a vector space. Production systems use learned embeddings; this lab uses a bag-of-words cosine so you can see overlap instantly.

Why can paraphrases score low here?

This toy model only counts shared words. Real embedding models can score “RAG looks up documents” and “retrieval-augmented generation” as similar even with few overlapping tokens.

Lab · Search & RAG · ~3 min

Embedding similarity

Compare two texts with a bag-of-words cosine score. See why embeddings beat keywords, then practice in the embeddings lesson.

How to read this

Use this to feel cosine similarity, not to pick a production embedding model.
Try a paraphrase with almost no shared words — the toy score drops; a real model often would not.
Continue in What are embeddings, then RAG.

Continue in What are embeddings? →All labs