An embedding is a piece of text converted into a list of numbers — a "vector" — arranged so that things with similar meaning end up close together. "Heart attack" and "myocardial infarction" use no words in common, but their embeddings sit almost on top of each other. That is the trick: embeddings let a computer compare meaning, not just spelling.
This is the quiet machinery behind modern search and the retrieval step in retrieval-augmented generation. When a tool "finds the most relevant passage," it is usually comparing embeddings, which is why it can surface the right paragraph even when your query and the source share no exact words.
Why it matters at your desk. For a researcher, embeddings are why Consensus and Perplexity can pull genuinely related studies and sources rather than keyword matches — semantic search finds the paper that answers your question, not just the one that repeats your phrasing.
What to watch for: "close in meaning" is the model's judgement, not a fact, and embeddings inherit the blind spots of the model that produced them. Search built on them can quietly miss a relevant document or rank a superficially-similar but irrelevant one highly — so for high-stakes work, treat semantic search as a strong first pass, not the final word on what exists.