SkillCDG: A Graph-Based Framework for Long SKILL Compliance Detection

A novel framework named SkillCDG has been introduced to enhance compliance detection within lengthy SKILL documents utilized in enterprise agent systems. As the complexity of business scenarios escalates, these documents have become increasingly prevalent. However, larger models result in high inference costs, while smaller models may struggle with accuracy in detection. SkillCDG structures intricate business policies using a two-layer constraint dependency graph: the upper layer organizes SKILL descriptions for scenario navigation, while the lower layer delineates dependencies among atomic constraints within each SKILL. Inference involves a two-tier retrieval process followed by dependency closure, facilitating compliance assessment and source traceability. Evaluations on three enterprise datasets and two controlled public benchmarks demonstrated that SkillCDG surpasses baseline models. The research can be found on arXiv with the identifier 2608.08146.