Knowledge graphs and ontologies are frequently proposed as essential components of GraphRAG architectures for grounding large language models, improving retrieval, and providing a semantic layer. However, discussions of these concepts often remain informal, with terms such as context, semantic layer, and even knowledge graph used without precise definitions. This ambiguity makes it difficult to compare approaches, evaluate implementations, or reason about their capabilities.
This session proposes a formal framework for defining context and the semantic layer in GraphRAG systems. Drawing on principles from knowledge representation and ontology engineering, this framework characterizes context as a structured, formally defined set of entities, relationships, constraints, and inferential mechanisms that govern the interpretation of retrieved information. The semantic layer then serves as the ontological foundation that enables consistent interpretation, integration, and reasoning across heterogeneous data sources.
We conclude by demonstrating how these formal definitions support the design and implementation of GraphRAG deployments, thus yielding systems that are more explainable, interoperable, and semantically robust than approaches based solely on unstructured retrieval.