It sounds like the metric you want is actually information content.
This may require a bit of hand-waving, since it isn't clear why "magic that requires you to understand the structure of a thing" would differentiate between not needing to care about the exact structure of a pile of randomly arranged atoms (e.g. a bucket of pure H₂O, which has extremely low information content) vs. say a snowflake which could be argued to have a much, much higher information content (i.e. to precisely describe its shape). But you said "pure" elements are supposed to be easy, so presumably this is what you want.
In this sense, although precisely specifying everything about a bucket of pure H₂O is easy. Yes, every single molecule has a unique position and velocity (and thanks to Heisenberg, you can't know both), but these properties are random. They aren't meaningful; they aren't information.
By this metric, something like a cut Emerald or a die-cast wrench is pretty easy; you only need to understand the basic structure of the atoms and their relative proportions, and have a fairly rough idea of the shape. A very simple integrated circuit would be very difficult, while something really complicated like a modern CPU or a living cell is almost certainly beyond possibility, at least if we're talking about humans. (You might make a virus or a DNA strand, but DNA is barely scratching the surface of the actual information that goes into living organisms. DNA is to living things what the recipe for stainless steel is to a modern skyscraper.)
Star Trek replicators are sometimes stated to work in this way, for much the same reasons; there is a huge difference in the quantity of information needed to describe the random arrangement of proteins and chemicals to make something that can pass for "steak", or even for a dead steer, and the quantity of information needed to create a living steer. Incidentally, this suggests an interesting wrinkle; your magic might be easier if you don't care about mangling your "source material".
(BTW, I've read books on this subject, though I couldn't recall specific titles offhand.)
To more directly answer your question, the complexity of a substance is proportional to how easy it is to describe it. Thus, the thing to do is to ask yourself how difficult it would be to describe a thing in a way that you would consider anything matching your description to be "the same thing". This is why something like "a bucket of water" is easy; a bunch of H₂O molecules in any arrangement is still "water". For something like a ball bearing, you need to know that it's a sphere, that it maybe has a certain crystal structure, that is has such-and-such proportions of iron, carbon, and so forth, but the exact positions of the carbon atoms isn't important. For something like a protein, well, that's a whole other kettle of fish... and forget something like a fish. As noted above, maybe you can wrap your head around it enough to have something that resembles a (dead) fish when you're done. OTOH, you might be able to do something like change all of the DNA in an organism to some other arrangement (we have, after all, examples of completely sequenced DNA), although what use this would be is a whole other matter.