What are embeddings?

Embeddings, i.e. vector representations, are lists of numbers that describe content in a way that's useful for comparing similarity. For a suitably chosen model, texts with similar meanings can have similar representations even when they use different words.

For example, the question “how do I request time off?” can be matched with a document called “leave request”. That doesn't mean similarity proves truth or currency. Two opposite statements about the same thing can be close in topic. Documents and queries must be encoded consistently; vectors from different models aren't automatically comparable just because they have the same number of values.