The index does not have to be tiny, the source code is. I tested with a 500k documents dataset.