Are embeddings only for documents? What about passages, sentences, words?
Do they work for documents of arbitrary length?
Depending on the embedding model they can work for words, sentences, passages, and any text up to ~2 pages of text (in the case of OpenAIs embedding model, most open source models can handle ~1 paragraph). If you want to go beyond that, and embed a document of 10 pages, you can average the embeddings of smaller sections (paragraphs/pages). Averaging in high dimensional spaces makes sense although different models handle it better than others