You pretty much summed up the drawbacks of the embeddings approach. In my experience it's pretty hard to extract the relevant parts of text, especially when the text is uniform.