Graphs RAG at Scale: Beyond Retrieval-Augmented Generation With Labeled Property Graphs and Resource Description Framework for Complex and Unknown Search Spaces

Researchers Manie Tadayon and Mayank Gupta have published a paper introducing 'Graph RAG,' a new framework designed to overcome the limitations of traditional Retrieval-Augmented Generation (RAG) systems. Traditional RAG struggles with unknown search spaces and semi-structured or structured documents, like databases or JSON files. The Graph RAG framework leverages two powerful graph architectures: Labeled Property Graphs (LPG) and the Resource Description Framework (RDF). This hybrid approach enables dynamic document retrieval without needing to pre-define the number of relevant sources, a major bottleneck in current systems.

A core innovation is a method for converting documents into RDF triplets using JSON key-value pairs, allowing semi-structured data to be seamlessly integrated into the knowledge graph. Furthermore, the team developed a 'text to Cypher' framework for querying LPGs, achieving over 90% accuracy in real-time translation of natural language questions into the Cypher query language used by graph databases like Neo4j. This makes complex, multi-hop reasoning queries fast and reliable for online applications.

The empirical evaluation presented in the paper shows that Graph RAG significantly outperforms traditional embedding-based RAG in accuracy, response quality, and reasoning, particularly for complex tasks involving interconnected data. By moving from a purely vector-based 'semantic search' model to a structured graph-based retrieval system, it can understand relationships and hierarchies within data that standard RAG misses. This positions Graph RAG as a transformative solution for the next generation of enterprise AI systems that need to reason over technical documentation, financial reports, or internal knowledge bases.