Implementation began from the NNDescent provided in the article blow, and end up by beeing an other code between the NNDescent and NNDescentFull, I didn't go further because of the number of concurency at the memory load, when computing the distance between vectors wich reduce the performence.

DONG, Wei, MOSES, Charikar, et LI, Kai. Efficient k-nearest neighbor graph construction for generic similarity measures. In : Proceedings of the 20th international conference on World wide web. ACM, 2011. p. 577-586.