Graphify.md replaces RAG and GraphRAG with pre-structured knowledge graphs that outperform on every structural query type — using 11× fewer tokens per query, with no hallucinations by construction. When domain graphs interact, intelligence compounds. That's CKGO.
An open benchmark across 45 domains, 7,928 queries, three retrieval architectures. Every result is reproducible.
| System | Macro F1 | Tokens/query | Run Cost | RDS |
|---|---|---|---|---|
| CKG | 0.471 | 269 | $7.81 | 0.00201 |
| RAG | 0.123 | 2,982 | $76.23 | 0.0000482 |
| GraphRAG | 0.120 | 3,450 | $44.43 | 0.0000452 |
Blue = CKG | Grey = RAG | T1 entity lookup is the designed negative control — CKG stores structure, not prose.
45 domains · 7,928 queries · three systems · all results and evaluation code published on GitHub. Co-authored with Dan McCreary (Intelligent Textbooks, ex-Optum). Pending ArXiv submission (cs.IR). github.com/Yarmoluk/ckg-benchmark →
Three steps from your data to a deployed knowledge system — no annotation budget, no expert curation required.
Any domain with stable relationships: regulatory registries, clinical trial databases, financial filings, product catalogues, policy documents, internal knowledge bases.
Entities, dependencies, and taxonomy are extracted into a Compact Knowledge Graph (CKG) — a directed acyclic graph encoding your domain's structure. Proprietary methodology. No hallucination by construction.
Queries resolve by traversing the graph — not by semantic similarity. Results are exact, reproducible, and hallucination-free by construction. 269 tokens per query average.
Track 2 built a GLP-1/Obesity pharmacology domain entirely from the ClinicalTrials.gov public API — no textbook, no domain expert, no annotation. The result exceeded the hand-curated educational average by 12.5%.
668 semaglutide trials · 224 tirzepatide trials · 158 pipeline agents (retatrutide, cagrisema, orforglipron). 90 concepts · 170 dependency edges · 170 benchmark queries.
vs. 0.471 hand-curated educational average — commercial domain outperforms expert-curated domains.
Expert annotation is not a prerequisite for the CKG advantage. Any domain with stable concept relationships expressible in a DAG achieves the same retrieval superiority.
The token efficiency holds on enterprise data: 11× fewer tokens per query, 28× compound RDS over RAG. CKG F1 = 0.530 vs. RAG 0.154 on the same queries.
Near-perfect enumeration of agents by drug class, indication by anatomy, and trial by program. RAG: 0.108. GraphRAG: 0.031.
Interactive — click nodes to explore dependencies · drag to reposition · Open full screen →
The GLP-1RA drug class node is the single highest-dependency hub in the graph — 20 concepts depend on it directly. Obesity pathophysiology (12×) and Weight loss endpoints (10×) form the next tier. Remove any of these three and the graph reorganizes structurally.
Every other drug in the DRUG taxonomy activates a single receptor pathway. Tirzepatide's simultaneous GLP-1 and GIP receptor activation is not a marketing claim — it is a structural position no other agent in the 90-concept graph shares. The graph made this unambiguous before reading a single trial paper.
The graph traces a 7-hop dependency chain from obesity pathophysiology through insulin resistance, visceral adiposity, metabolic syndrome, dyslipidemia, cardiovascular disease, and MACE endpoints. The SELECT trial outcome was visible in this architecture before the data published — the path existed in the structure.
SUSTAIN, STEP, SURMOUNT, AWARD, LEADER, SCALE, SELECT, CVOT design — all 14 programs mapped with their endpoint dependencies. Seven taxonomies (FOUND, PATH, DRUG, TRIAL, COMPL, SPEC, COMBO) partition 90 concepts into reasoning layers an LLM can traverse without hallucination.
27 verticals deployed. The architecture is identical across domains — only the knowledge graph changes.
Live interactive knowledge graphs — any topic, production-ready in minutes.
Patent-protected methodology, peer-reviewed benchmark, and a construction pipeline that scales to any structured domain.
Patent Pending · USPTO · Priority date locked
April 16, 2026 · Provisional application on file
Conversion in progress · Represented by patent counsel
Microsoft, Intel, Unlikely AI · Citations: Buehler Lab · MIT
44 domains · 7,758 queries · F1 3.7× over RAG · 42× RDS · Published open benchmark — academic citation trail established before non-provisional filing.
GLP-1/Obesity Track 2: pipeline-generated CKG from ClinicalTrials.gov API, zero human annotation. F1 0.53 — exceeds hand-curated average by 12.5%. The method scales to any structured domain automatically.
RAG searches by similarity. It doesn't understand structure. Accuracy, zero hallucination, and 80% token reduction aren't aspirational — they're the result of searching structure, not probability.
RAG guesses from all available information. CKG traverses pre-built dependency paths — no vector similarity, no hallucination. The answer is in the graph or it isn't there.
When structured domain graphs interact, emergent connections appear that no single graph — and no retrieval pipeline — could surface. This is CKGO: the orchestration of knowledge, not agents.
As AI replaces search, the question isn't how to rank your content — it's whether your domain knowledge is structured enough for AI to cite accurately. CKG is the answer layer for GAO.
The technology is built. The benchmark is published. The patent is filed. The next step is a 30-minute call to scope your domain and structure a pilot.
Schedule a 30-Minute Call
Daniel Yarmoluk · Founder · Graphify.md
daniel.yarmoluk@gmail.com