Autonomous robotic surgery has advanced significantly based on analysis of visual and temporal cues in surgical workflow, but relational cues from domain knowledge remain under investigation. Complex relations in surgical annotations can be divided into intra- and inter-relations, both valuable to autonomous systems to comprehend surgical workflows. Intra- and inter-relations describe the relevance of various categories within a particular annotation type and the relevance of different annotation types, respectively. This paper aims to systematically investigate the importance of relational cues in surgery. First, we contribute the RLLS12M dataset, a large-scale collection of robotic left lateral sectionectomy (RLLS), by curating 50 videos of 50 patients operated by 5 surgeons and annotating a hierarchical workflow, which consists of 3 inter- and 6 intra-relations, 6 steps, 15 tasks, and 38 activities represented as the triplet of 11 instruments, 8 actions, and 16 objects, totaling 2,113,510 video frames and 12,681,060 annotation entities. Correspondingly, we propose a multi-relation purification hybrid network (MURPHY), which aptly incorporates novel relation modules to augment the feature representation by purifying relational features using the intra- and inter-relations embodied in annotations. The intra-relation module leverages a R-GCN to implant visual features in different graph relations, which are aggregated using a targeted relation purification with affinity information measuring label consistency and feature similarity. The inter-relation module is motivated by attention mechanisms to regularize the influence of relational features based on the hierarchy of annotation types from the domain knowledge. Extensive experimental results on the curated RLLS dataset confirm the effectiveness of our approach, demonstrating that relations matter in surgical workflow analysis.