Knowledge discovery presupposes mining the inherent relationships between the numerous data records and categories within a given database. Clustering is one of the popular approaches to such data mining. Most of these clustering techniques require an a priori knowledge of the intrinsic number of clusters or groups in the database, and in addition, their interpretation needs some externally definable basis of affinity underlying each of the clusters (cluster class labels). However, in a real-world knowledge discovery process, such a priori knowledge is not often available, and the user is interested in discovering this as well. In this study, a new approach aimed at discovering the intrinsic number of groups, at the most elemental level, which may then be combined to form metagroups to the extent desired, is proposed. This approach is based upon a concept, referred to here as reciprocal relationship bonds. The method initially identifies all data record pairs in the database with such reciprocal relationship bonds. These represent the cores of the data record groups, to be formed by step-wise bonding of all the records in the entire database. Various levels of relationship can then be defined between any pair of records, depending on the number of bonds required to connect these data records. The strength of bond of each record to the cluster can be ordered, based on how far removed it is from the cluster core. The lower the order, the stronger is its bond to the cluster, and higher is the likelihood that the data record truly belongs to the corresponding cluster. The approach also provides a mechanism for flagging the out-of-norm or unusual data, by clustering them separately from other normal data records, which may indicate incomplete or error-prone records.