Embeddings - the math of meaning
An embedding is a way of converting text into a list of numbers that captures its meaning, so that similar meanings end up as similar numbers - even if the actual words are totally different. This is what lets a search system find "vehicle maintenance costs" as relevant to a query about "truck repair expenses," despite no shared words.
Text gets passed through an embedding model, producing a vector (a long list of numbers) representing its position in a huge "meaning space." Searching then becomes a math problem: find the stored vectors closest to your query's vector. This is the retrieval half of RAG (C1) and the core of any "semantic search" feature.
No task needed to grasp this conceptually - but if curious, search "embedding projector visualization" to see an actual 3D map of words/sentences clustered by meaning; seeing similar concepts cluster together makes the idea concrete fast. Popular embedding models you'll see named in the wild: OpenAI's text-embedding-3, Voyage AI (Anthropic's recommended partner), and Cohere Embed.
What does an embedding actually capture about a piece of text?