The exponential growth of digital products lead to the massive generation of data every day at the alarming volume, velocity and diversity. In order to extract meaningful value, advanced storage ,processing and analytical systems are growing fast. Collective Adaptive Systems(CAS) are such large scale distributed systems which provide a scalable and efficient way of storing and processing the data at such a massive scale. Allow Ensemble[13] is one such CAS systems where participants(nodes) have the knowledge repository based on probabilistic graphical models. These model are built from the observed data in different contexts. These models are associated with a degree of learning based on the quality and amount of observed data. The participants of CAS are loosely connected components differing in dimensionality of these knowledge repositories in a dynamic environment. The efficient retrieval of knowledge from these probabilistic models possess a greater challenge and is the core task of this thesis work. The major challenges we faced to achieve this task includes: 1.A confidence measure to quantify the degree of learning of the knowledge model on each participant (network node). 2.This confidence measure should answer the query in absolute sense in a way that it determine the confidence in knowledge model wrt to itself when tuned to give the best/saturated knowledge. 3.The mechanism needs to be developed to aggregate this measure for a group of crf nodes to determine the overall average degree of learning. 4.The efficient routing mechanism need to be developed to answer the query with specified confidence measure of learning. We propose the following concepts to handle this task. 1.Confidence Metric which quantify the degree of learning and providing the absolute sense of learning using the statistical concept of Confidence intervals. The confidence value is associated with every node of the probabilistic crf graph and these values can be aggregated to produce an overall confidence score. 2.We have developed an efficient routing mechanism so that query can be routed to the best learned knowledge model specific to its context parameters. Each participant stores and propagates the information about the knowledge model reachable through its neighbor. We call this approach knowledge aggregation based routing. 3.We have proposed a new routing scheme of query learning in which routing behavior is developed from the past queries. In general, the high retrieval accuracy is obtained in both scenarios. We believe that our approach of summarization ,propagation and effective retrieval of knowledge in a probabilistic models poses great direction to further enhance and develop interesting applications that can make good use of our research work.