Search Results

Retrieval Augmented Generation (RAG) is seeing rapid adoption in industry to enable employees to query information captured in proprietary data for their organisation. In this work, we test the impact of domain-relevant knowledge graph schemas on the results of Microsoft’s GraphRAG pipeline. Our approach aims to address the poor quality of GraphRAG responses on technical reports rich in domain-specific terms. The use case involves technical reports about geology, chemistry and mineral processing published by the Minerals Research Institute of Western Australia (MRIWA). Four schemas are considered: a simple five-class minerals domain expert-developed schema, an expanded minerals domain schema, the Microsoft GraphRAG auto-generated schema, and a schema-less GraphRAG. These are compared to a conventional baseline RAG. Performance is evaluated using a scoring approach that accounts for the mix of correct, incorrect, additional, and missing content in RAG responses. The results show that the simple five-class minerals domain schema extracts approximately 10% more entities from the MRIWA reports than the other schema options. Additionally, both the five-class and the expanded eight-class minerals domain schemas produce the most factually correct answers and the fewest hallucinations. We attribute this to the minerals-specific schemas extracting more relevant, domain-specific information during the Indexing stage. As a result, the Query stage’s context window includes more high-value content. This contributes to the observed improvement in answer quality compared to the other pipelines. In contrast, pipelines with fewer domain-related entities in the KG retrieve less valuable information, leaving more room for irrelevant content in the context window. Baseline RAG responses were typically shorter, less complete, and contained more hallucinations compared to our GraphRAG pipelines. We provide a complete set of resources at https://github.com/nlp-tlp/GraphRAG-on-Minerals-Domain/tree/main. These resources include links to the MRIWA reports, a set of questions (from simple to challenging) along with domain-expert curated answers, schemas, and evaluations of the pipelines.