Inner product distance is not normalized. Cosine distance is actually -1 to 1 (0 being neural) which we scale back to 0 to 2 because that's the convention.