The first step is to create a grid of vectors, where each vector points in a random direction. A vector is basically an arrow with a magnitude (length) and a direction(angle).
The important thing about these vectors is that they must all have a magnitude of 1. This is what we call a ‘normalised’, or 'unit-length' vector, where the length has a magnitude of 1.
In our example of generating a 2D heightmap, you can imagine the vectors as arrows with a length of 1, pointing in any direction in a circle (360 degrees). This grid must be a lower resolution than the final perlin noise wanted. For example if you wanted some Perlin Noise to fill a 1000x1000 grid, you might only have a 50x50 grid of vectors. This ratio would be adjusted depending on the use case.