Introduction

Motivation

LLM retrieval quality is typically measured by F1 alone, yet token cost is a first-class production constraint---not a research afterthought. Domain-specific knowledge has latent structure that RAG discards through chunking, and GraphRAG re-derives structure from text at significant computational expense. But what if the structure is already known?

The Three Paradigms

The interactive diagram below illustrates the three retrieval pipelines compared in this benchmark. Hover over any component to learn more about its role.

System	Knowledge Representation	Retrieval Method	Build Cost
RAG	Unstructured text chunks	Embedding similarity	Embed all chunks
GraphRAG	Dynamically extracted graph	Graph + community search	Full entity extraction
CKG	Pre-structured DAG + taxonomy	Direct concept/edge lookup	Zero (CSV-native)

Falsifiable Claims

We make the following falsifiable claims, each testable against the benchmark results:

1. CKG achieves higher F1 on T2 (dependency) and T3 (multi-hop path) queries.
2. CKG F1 does not degrade with hop depth; RAG F1 degrades significantly.
3. CKG RDS ratio \(\geq\) 10x vs RAG across all 25 domains.
4. GraphRAG hallucinates edges not present in ground truth DAG (HR > 0).
5. CKG Hallucination Rate = 0 (by construction).
6. The Structure Premium hypothesis: RDS advantage correlates with DAG richness (\(r > 0.7\)).

Contributions

1. The CKG architecture specification (format, DAG constraints, taxonomy schema).
2. Five novel evaluation metrics (RDS, CUR, Hop-F1, CPCA, RP).
3. The McCreary Corpus as a formal benchmark dataset (first paper to do so).
4. An open benchmark: 25 domains x ~175 queries x 3 systems (~13,000 evaluated pairs).
5. A HuggingFace dataset with one-command reproduction harness.