HCG-RAG Improves RAG Quality with Compact Causal Graphs

Graph-based Retrieval-Augmented Generation (GraphRAG) systems typically extract entities and relationships exhaustively, leading to knowledge graphs that scale in size and construction cost with the length of the source corpus. This often results in overly large and expensive graphs that are not necessarily optimized for the specific reasoning a query requires. HCG-RAG proposes a more efficient approach by replacing open-ended extraction with schema-constrained causal graphs. This automated pipeline distills a corpus into a fixed, typed vocabulary of causal variables and constructs a compact, two-tier graph. This method drastically reduces the number of nodes and LLM calls during graph construction, making the graphs auditable by domain experts. On medical and clinical benchmarks, HCG-RAG matched or surpassed leading entity-relation systems, demonstrating that structured, higher-level organization within the graph is more critical than sheer scale for effective retrieval.