ARTFEED — Contemporary Art Intelligence

HCG-RAG: Schema-Constrained Causal Graphs for Efficient GraphRAG

ai-technology · 2026-07-29

A recent paper published on arXiv presents HCG-RAG (Hierarchical Causal Graph RAG), a novel technique that substitutes the exhaustive entity extraction process in GraphRAG with causal graphs constrained by schema. This system employs an automated pipeline to refine a corpus into a fixed set of causal variables, resulting in a streamlined two-tier graph. HCG-RAG demonstrates a significant reduction in node usage, requiring 3-20 times fewer nodes than baseline methods, and decreases LLM calls by a factor of 8x-135x, all while maintaining answer quality. Evaluations on medical and clinical benchmarks, including a validated epilepsy dataset by neurologists, indicate strong performance. Additionally, this method allows domain experts to review and amend the graphs.

Key facts

  • HCG-RAG uses schema-constrained causal graphs instead of exhaustive entity extraction.
  • Graph size and build cost are decoupled from corpus length.
  • 3-20x fewer nodes than entity-relation baselines.
  • 8x-135x fewer LLM calls than MS-GraphRAG.
  • Validated on medical and clinical benchmarks including an epilepsy dataset.
  • Graphs are compact enough for domain expert audit and correction.
  • Published on arXiv with ID 2607.22592.
  • The method matches answer quality of entity-relation baselines.

Entities

Institutions

  • arXiv

Sources