Abstract

Retrieval-augmented generation (RAG) and graph-based retrieval (GraphRAG) are the dominant paradigms for grounding LLM responses in structured knowledge. Both optimize for retrieval recall while treating token cost as a secondary concern. We introduce Compressed Knowledge Graphs (CKG)---pre-structured DAG representations with explicit concept taxonomy and pipe-delimited dependency encoding---and present the first large-scale benchmark comparing all three architectures across 46 domains.

We evaluate on the McCreary Intelligent Textbook Corpus: 46 open-source educational domains (calculus, chemistry, computer science, genetics, bioinformatics, economics, machine learning, and more), each with a standardized learning graph CSV totaling 12,260 concepts and 19,405 dependency edges. We introduce five novel metrics: Reasoning Density Score (RDS = F1 / tokens), Context Utilization Rate, Hop-Depth F1 Degradation, Cost Per Correct Answer, and Relationship Precision.

Results Pending

Experimental results will be added after benchmark runs are complete.

Benchmark, dataset, and evaluation harness are released at HuggingFace under CC BY 4.0 / MIT.

ArXiv categories: cs.IR (primary), cs.AI (secondary)