the simple answer is that the datamart/datawarehouse must contain raw detailed records, otherwise you can't create the OLAP cube

the OLAP cube should only contain summarised data, otherwise they become slow to build

the whole point of an OLAP cube is that it is pre-summarised data created during the daily build from the datamart/datawarehouse

the ETL process then becomes

OLTP --> staging --> datamart/datawarehouse --> OLAP cube

when writing reports you use the datamart/datawarehouse for detailed ones, e.g. show me all the line items ordered by a particular customer

and you use the OLAP cube for summaries, e.g. total sales by city for the last six months

see http://www.kimballgroup.com/ for advice on datawarehousing

regards

Andrew