As long as performance isn't a concern, you can do whatever's convenient - which probably means a separate vertex/index buffer for each mesh.
If you sort your meshes when drawing, so that all instances of the same mesh are drawn together (e.g. all units of type A are drawn, then all units of type B, etc.), you can reduce state changes and also set yourself up well for adding geometry instancing later if it becomes necessary for performance. This will be a good pattern if you don't have too many different meshes, but potentially many instances of each mesh - which seems likely for an RTS (not too many unit/terrain types, but potentially a lot of units/tiles on screen).
The "one huge buffer" approach can reduce the number of state changes / draw calls still further if necessary, but it takes more code to manage, and requires a few other things like texture atlasing in order to actually realize the performance gains. I wouldn't worry about it unless and until performance analysis indicates that you truly need it.