This is the kind of problem treated in finite population sampling theory, as presented in the book http://www.amazon.com/Finite-Population-Sampling-Inference-Prediction/dp/0471293415/ref=sr_1_1?s=books&ie=UTF8&qid=1401276486&sr=1-1&keywords=finite+population+sampling+theory
(and many others).
First of all, you will want a practical way to do the sampling! So we will need to know how your physical files are stored! You should also think about if they are stored in completely random order, or in some other definite order (alphabetical? chronological? what???). A simple random sampling will often be impractical --- could be practical if you have some list of all the files in your collection, wheather that list is on a computer or on paper (and if it is practical to access the files out of order). If the answer to that question is NO, so simple random sampling is impractical, you can use some kind of cluster or stratified sampling. You can find explanations of this and related terms here: https://en.wikipedia.org/wiki/Statistical_sampling
Without knowing more about the practical situation, only some brief hints: sampling could be on the level of
or some other units in which the physical storage is organized. You should think about questions such as: If the storage is is say, chronological order, might there be some trend in document size with time? Or some cyclical variation, that is , in some parts of the year, documents of a specific type are produced, which typically differs in length?
We can help more if you tell us more about the situation!
(Then, depending on the sampling plan choosen, there will be some specific formulas/methods to use to construct estimates and confidence intervals)