Graph-guard, a knowledge graph behind RAG for multi-hop retrieval (github.com)

A new tool called Graph-guard has been introduced to enhance Retrieval-Augmented Generation (RAG) by utilizing a knowledge graph for multi-hop information retrieval. Traditional RAG models struggle with complex queries that require synthesizing information from multiple documents, as they typically rely on direct text similarity. Graph-guard addresses this challenge by implementing a typed knowledge graph that can trace connections between facts, resulting in a 14% improvement in hit rates and a 26% boost in mean reciprocal rank (MRR) for multi-hop questions, while maintaining performance on straightforward lookups. What sets Graph-guard apart is its two-tiered system, with Tier A employing a lean knowledge graph for fast retrieval and Tier B incorporating a full semantic layer using RDF, OWL, and SPARQL for richer data validation and standard compliance. However, the findings indicate that while the ontology is beneficial for data fidelity and interoperability, it does not significantly enhance retrieval performance. This nuanced understanding of when and how to apply complex semantic structures is pivotal for the AI/ML community, allowing practitioners to optimize their systems based on specific retrieval challenges rather than defaulting to heavier computational methods.