Composing learning material for online education and ensuring quality of delivery (low latency and high bandwidth) is a challenge to researchers in e-learning. The problem is compounded when the computing platforms are heterogeneous. Besides, e-learning as an application presents a highly fluctuating load to the underlying IP Multimedia Subsystem. Viewed in this light, it appears that an e-learning system with inherent abilities to integrate various multimedia elements that can adapt its output to suit network parameters is the need of the hour. In this paper, we propose a new architecture for the e-learning application - called ACCS - Adaptive Content Composition System - which is capable of composing a knowledge molecule (composed of different media elements) that can parametrically relate to the underlying IP Multimedia Subsystem (IMS), for ensuring a desired QoS. The ACCS architecture has two dominant layers above the IMS; viz., knowledge management sub-layer and network services adaptation sub-layer. We study the performance of ACCS with synthetic workloads under controlled conditions using a network test-bed, called GeNIe.