Retrieval-Augmented Generation (RAG) was a meaningful step forward, but it introduced a new set of failure modes that compound at scale. CKG was designed to fix all of them.
The core insight: RAG's problem isn't retrieval speed — it's retrieval quality. Cheap tokens don't fix bad data. Structure wins over volume.
Across a reproducible benchmark of 45 domains and 7,928 queries, CKG used a mean of 269 tokens per query compared to 2,982 for RAG — an 11× reduction.
Source: Yarmoluk & McCreary, "Compact Knowledge Graphs vs. RAG and GraphRAG: A Reproducible Benchmark Across 45 Educational Domains," arXiv 2026. Benchmark: 12,261 nodes · 19,626 edges · fully reproducible. Full benchmark on GitHub →
If your team runs 10,000 LLM queries per month against a domain knowledge base, the difference between RAG and CKG is not academic:
The accuracy improvement isn't magic — it's structural. When an LLM receives pre-structured knowledge with explicit entity relationships, it doesn't have to guess.
Retrieve the right chunks, then reason over noisy, unstructured text. Each step compounds error. The model hallucinates when the retrieved context is ambiguous, incomplete, or contradictory — which it often is.
A CKG pre-encodes entities, relationships, and dependencies before the query runs. The model receives a structured map of the domain, not a pile of text chunks. It reads the graph rather than inferring it.
Example (GLP-1 payer coverage): A RAG system retrieves 12 formulary PDF chunks and asks the model to determine whether Ozempic is covered at Tier 2 for a Type 2 diabetes indication under a specific Medicare Advantage plan. The CKG encodes Drug → Payer → Plan → Tier → Indication → Prior Auth requirement as explicit relationships. The model reads the answer directly. F1: 0.5306.
The CKG format uses typed relationships, dependency declarations, and taxonomy labels. There is no ambiguity for the model to resolve — and therefore no hallucination surface.
Retrieval Density Score (RDS) is the primary metric for measuring knowledge graph efficiency. It quantifies how much correct information you receive per token spent.
RDS = F1 Score / Mean Tokens Used CKG: 0.4709 / 269 = 0.001751 RAG: 0.1231 / 2982 = 0.0000413 CKG RDS advantage: 42×
A higher RDS means your LLM is getting more accurate answers for less money. RDS penalizes both inaccuracy and token bloat — a system that is accurate but verbose scores lower than a system that is accurate and compact.
Graphify.md introduced RDS as a standardized benchmark metric for comparing knowledge delivery systems. It is included in the published arXiv benchmark paper.
| Approach | Token Cost | Accuracy | Infrastructure | Domain Updates |
|---|---|---|---|---|
| Compact Knowledge Graph | 269 tokens avg | F1: 0.4709 | None — one .md file | Swap the file |
| RAG (vector retrieval) | 2,982 tokens avg | F1: 0.1231 | Vector DB + embeddings | Re-embed changed docs |
| Fine-tuning | Minimal at inference | Domain-dependent | GPU cluster + data pipeline | Retrain for every update |
| Graph database (Neo4j, TigerGraph) | Low per query | High if schema correct | Graph DB + Cypher + API layer | Schema migrations required |
| Unstructured context stuffing | Unpredictable (high) | Low — noise dominant | None | Paste new text |
CKG is the only approach that combines zero infrastructure overhead with high accuracy and low token cost. It is not a retrieval system — it is a pre-structured context format.
Traditional knowledge graphs (Neo4j, TigerGraph, AWS Neptune) are databases. They require:
A Compact Knowledge Graph is a serialized text file. It encodes the same entity relationships in a format LLMs can read natively — no query language, no database, no pipeline.
ConceptID,ConceptLabel,Dependencies,TaxonomyID 1,GLP-1 Receptor Agonist,,FOUND 2,Semaglutide,1,CORE 3,Ozempic (Brand),2,CORE 4,Medicare Advantage,1,CORE 5,Tier 2 Formulary Coverage,3|4,ADV 6,Type 2 Diabetes Indication,2,CORE 7,Prior Authorization Required,5|6,ADV
Drop this into your LLM system prompt. The model reads entity IDs, labels, dependency chains, and taxonomy tags — and answers questions about formulary coverage, prior auth requirements, and drug-payer relationships without retrieving a single document.
CKGs are highest-value in domains with structured, high-stakes, frequently-updated information that is sparse in LLM training data.
Graphify.md builds CKGs from public data sources using a proprietary compression pipeline. The output is a pair of files:
Public sources: SEC EDGAR, USPTO, GDELT, USASpending, openFDA, ClinicalTrials.gov, and domain-specific repositories. Each vertical draws from the sources most relevant to its entity structure.
GitHub repository → raw file URL → API-accessible JSON. No infra required on the customer side. Weekly update cadence for live-data domains.
Production benchmark: 27 verticals deployed in 60 days. 12,261 nodes · 19,626 edges · 45 domains benchmarked. One operator.
No — CKG accelerates everything you've already built. It is not a platform, a database, or a framework. It is pre-structured domain knowledge that makes every layer of your AI stack perform better.
The positioning: Graphify.md is not competing with your AI infrastructure investment — it is the knowledge layer that makes that investment pay off. Think of it as the domain expertise your AI was missing.
Graphify.md is the company that builds and delivers Compact Knowledge Graphs at scale. Founded by Daniel Yarmoluk (St. Louis Park, MN), Graphify.md operates a multi-domain CKG production environment that deploys across 27 verticals simultaneously.
The benchmark methodology and RDS metric were introduced in a peer-reviewed arXiv paper co-authored with Dan McCreary (former Senior Distinguished Engineer, UnitedHealth Group; patent holder US 11,204,950).
Scientific foundation includes citations from Markus Buehler (MIT) on cross-domain knowledge graph emergence and scale-free network architecture.
Tell us the domain. We'll show you what a CKG looks like for your specific use case — in one session.
Book a 30-Minute Demo Learn More →