when you say long context models as less effective for documents that consist of short sentences, do you mean that embedding models that have long context capabilities tend to be worse with shorter sentences or are you just saying that _using_ their large context windows will be less effective for docs with short sentences