You could assign the cluster based on what the k nearest neighbors are, if there is a clear majority. The quality will depend on the suitability of your embeddings.