Inverted-index posting-list evaluation

Design the low-level posting-list storage and query-evaluation engine of an inverted index that must answer boolean + phrase queries over a 5B-document corpus with very long posting lists for common terms (the term 'the' appears in billions of docs). A naive 'intersect two sorted doc-id lists' is too slow when one list is billions long and the other is short. How do you store posting lists compactly and evaluate AND/OR/phrase queries so the cost scales with the smaller list and with the result, not with the giant list?