It's very simple.
You have two entities: BaseItem and Item. Item derives from BaseItem. This means that if you tell the fetch logic to fetch BaseItem instances, you will get Item instances as well, because Item instances are also a BaseItem.
This is how inheritance on a relational model is implemented. The thing is that every BaseItem row is either a BaseItem OR part of an Item, as Item's data is spread across 2 entities: BaseItem and Item.
So if you tell the fetch logic to get all BaseItem instances, you WILL get data belonging to Item, and thus this data is stored in an Item instance instead, loading the Item data as well.
If you JUST want to load BaseItem instances which AREN'T an Item, you therefore need all rows which aren't part of an Item and thus which aren't of type Item.
If you want ALL BaseItem rows, no matter if they're in a BaseItem entity or in an Item entity, you can use a projection fetch: fetch BaseItem fields onto a BaseItem entity collection.
Say I have an inheritance model like Employee <- Manager <- Boardmember. (bad model, but illustrates the point). Say I fetch all Employees. What do you expect to get? All employees with just their Employee data? Or all Employees in their own type instance? (e.g. a manager has his data loaded in a ManagerEntity instance etc.)
I would definitely expect the latter, because if I want to load the data of manager because I want to work on manager, I have to perform per type a new query manually.
This all comes from the fact that the data of a subtype is fragmented across all supertype tables plus the table of the subtype. So when you look at a row in one of the tables, this is a fragment of an entity, and when you load entities, you can't deal with fragments, you need all data.
There are other things to consider: you might want to look into destroying the inheritance hierarchy between baseitem and item here, and just keep the 1:1 relation. This is perfectly workable.
In v2.1, we'll add delay loading for LOB fields (text/image/blob/clob etc.) so you can fetch them later on into an existing graph. This could solve your problem so we're aware of this and will offer a solution, yet that's not helping you today.
I hope this clears things up. If not, please feel free to ask more