That graph doesn't make much sense. The Y axis reads Relative Responsivity and gives a unit of µW/cm², which is not a relative unit.
It is probably the count value they obtained with some fixed irradiance and varying wavelength for the different sensors. This doesn't help much without knowing their test setup, which isn't given in the application note.
So I would just ignore this and work with the datasheet, which is still a bit of a mess, but gives better values to work with.
First we have the irradiance response given in counts per µW/cm² for different wavelengths:
Which is closer to what you need. But then they chose the wavelengths to measure this not at the peak responsivity. Which are:
Which does not correspond to the plot they give later on:
So you can now scale the curves according to the given points in the table. Which would make the peak of red read something like 96/0.7 = 137 counts/(µW/cm²), the peak of green 74/0.5 = 148 counts/(µW/cm²) and the peak of blue 56/0.9 = 62 counts/(µW/cm²). For accurate numbers use Engauge Digitizer.
I have no idea what is going on with white. It's almost non existent in the datasheet and the normalized response shows it being very sensitive in IR, which at least I cannot see and I usually don't count as white...