Building pre-structured knowledge graphs that outperform RAG at 11× lower token cost.
Founder of Graphify.md — CKG architecture, benchmark design, and enterprise deployment.
42× more intelligence per token than RAG. Zero hallucinations by construction.
45 domains · 7,928 queries · Fully reproducible
| System | Macro F1 | Tokens/query | RDS |
|---|---|---|---|
| CKG | 0.4709 | 269 | 0.00175 |
| RAG | 0.1231 | 2,982 | 0.0000413 |
| GraphRAG | 0.1200 | 3,450 | 0.0000452 |
CKG F1 improves with hop depth (0.374 → 0.772 at hop=5). RAG plateaus at hop=2.
→ ckg-benchmark — benchmark repo, paper, reproducible results
Pre-structured knowledge as a plain .md or .csv file. Drop it in your LLM context. No graph database, no embeddings, no retrieval pipeline.
The structure is the signal — not the curation effort. Track 2 proved this: a GLP-1/Obesity pharmacology CKG built entirely from the ClinicalTrials.gov API in one session achieved F1 = 0.530, exceeding the hand-curated educational average.
| Project | What it is |
|---|---|
| ckg-benchmark | 45-domain RAG vs CKG vs GraphRAG benchmark — paper + data + code |
ckg-mcp |
MCP server — CKG as a pre-compiled routing layer for agent stacks (coming) |
| Graphify.md | Commercial CKG deployment — weekly-updated domain knowledge for enterprise AI |
Benchmarking Knowledge Retrieval Architectures Across Educational and Commercial Domains: RAG, GraphRAG, and Compact Knowledge Graphs Yarmoluk & McCreary, 2026 · v0.6.2 · Patent pending App #64/040,804
→ Read the paper · PDF