A scalable, effective, and general way to augment LLMs with billion-scale knowledge graphs using very little GPU memory cost.