I think a pure diff is going to give you too much noise and will be misled by simple mutations (I think..). Here's an approach - after you normalize the docs t

I think a pure diff is going to give you too much noise and will be misled by simple mutations (I think..). Here's an approach - after you normalize the docs to text you should then extract terms/keywords (google "term extraction" - there are libs that use POS to help) and save those as tokens. One approach would be to use something like elasticsearch to do a "more like this" filter. The idea is that you want to find documents that are most similar to a certain document. I haven't tried this approach but if it's effective, it should be able to handle a million resumes easily.