Open Compressed Knowledge Graph standard · Patent-pending engine

Own the knowledge layer. Rent the model.

The problem was never the AI or your knowledge — it's the space between them. Graphify.md compresses your domain into a typed graph the agent reads before it acts: ~269 tokens, portable across any model, owned by you.

54 live, interactive CKGs · 12,978 nodes · open, reproducible benchmark · runs on a local 8B — no cloud.

Open
format · any model
269
tokens/query · vs 2,982
0.471
Macro-F1 · vs 0.123 RAG
~10×
cheaper to run · $7.81 vs $76.23
0%
hallucination by construction

CKG Explorer — live, interactive, in your browser

CKG Explorer — interactive graph traversal demo

Try it live → Free CKG Library · 61 domains · drag to explore · no signup

One platform, two halves. The open standard — format, client, free CKGs, benchmark — is yours to adopt at no cost. The engine that turns your corpus into a faithful, current graph is what we sell. New here? Start with See it run, or explore the live CKG Library.

For investors

Own the layer the frontier can’t reach.

Every model release torches a layer of startups built on the model. Graphify.md sits in the one place the next release can’t kill — the knowledge layer above it.

The bet

Own the knowledge layer. Rent the model.

An 8B model on a laptop, reading a Graphify.md CKG, scored 0.471 Macro-F1 — and climbed to 0.772 on 5-hop questions where RAG sits at 0.170. The graph does the hard part; the model just reads it. Swap the engine next quarter — the owned, portable knowledge layer doesn’t move.

The moat

The method, not the model.

Three provisional patents on automated structure discovery + compression, and an open, reproducible benchmark anyone can clone — 7,928 queries, 45 domains. We don’t out-compute the frontier; we make it cheaper to rent.

The stage

Early — by design.

Pre-revenue pilots, one open benchmark, a category still being named. Active conversations across enterprise data, supply chain, and compliance. That’s the entry point.

Book 20 minutes →

Why now — the tailwinds we sit on

The model layer is commoditizing. Value lodges in the knowledge layer above it.

Small-model proliferation

Open, small models are everywhere and up to ~100× cheaper to run — but only as smart as the knowledge they reason over. We're that layer: we make an 8B punch above the frontier.

Sovereignty

~95% of execs call owned / sovereign AI mission-critical. The knowledge layer is the asset they own — not rented from a frontier lab.

Deplatform risk · access rights

Models get deprecated, nerfed, and pulled — a frontier model was force-disabled worldwide in June 2026. An owned, portable layer is the backstop no external directive can switch off, and no one can blackball.

Cost & token efficiency

As $/token falls, value moves to whatever decides which tokens you needed. 11× fewer tokens, ~42× retrieval density — the cheapest correct answer.

Best scores

~4× the accuracy of RAG, multi-hop that rises with complexity, 0% hallucination by construction — open, reproducible benchmark.

Environmental · per-watt

Fewer tokens = less compute = less energy per correct answer. Cognition per watt — as AI's power draw becomes a board-level line item.

Plus an open standard (adoption flywheel — free CKGs via ckg-mcp) and a knowledge layer that stands up in hours, not quarters. Cheapest, best, fastest, owned — and deplatform-proof.

Knowledge architecture · vs the LLM Wiki

The LLM Wiki beats RAG. A Compressed Knowledge Graph beats the wiki.

The field just agreed a folder of plain-text files outperforms a vector database. They're right — and that's the floor, not the ceiling. The rung above it: a typed, deduplicated, provenance-tagged graph your agent traverses deterministically.

The point: deterministic beats probabilistic, readable beats opaque, density beats dump. The LLM Wiki proved all three against RAG. A Compressed Knowledge Graph is what happens when you engineer the wiki instead of hand-piling it.

One ladder, three rungs

Not three competing camps — three rungs of one staircase, each fixing the failure mode below it.

Rung 1 · RAG

The vector database

Chunk, embed, similarity-search, hope. Right for a 10M-document archive; overkill and opaque for everything smaller. You can't read it, audit it, or debug a wrong answer.

Rung 2 · LLM Wiki

The flat folder of files

Plain Markdown an agent loads selectively — deterministic at the file level, readable, Git-versionable. A real step up. But it's a pile: links are implicit, facts repeat, nothing's deduplicated, and "load the whole file" wastes tokens as it grows.

Rung 3 · Compressed Knowledge Graph

The engineered layer

The wiki's instinct, engineered: typed nodes and edges, deduplicated, dependency-ordered, every claim provenance-tagged. The agent traverses to the exact sub-graph — node-and-edge precision, not whole-file loading — and you can still read every line. Deterministic and dense.

What breaks in a flat wiki

We like the pattern enough to be honest about its ceiling. Four things degrade as a folder of files grows:

  • Redundancy compounds — the same fact in five files; the agent loads all five.
  • Relationships stay implicit — a [[link]] says two pages relate, not how (depends-on, contradicts, supersedes, causes). The reasoning lives in edges a flat wiki doesn't have.
  • Retrieval is whole-file — you wanted one fact; you loaded the whole page. Fine at 20 files, expensive at 500.
  • No provenance, no audit — which source backs this line? For regulated work, "the agent read it somewhere" isn't an answer.
The wiki asks an LLM to re-derive structure from prose every time it reads. A CKG hands the structure over pre-built — so the model spends its tokens reasoning, not re-parsing.

RAG vs. LLM Wiki vs. CKG

The popular comparison stops at two columns. Here's the third.

FactorRAGLLM WikiCompressed Knowledge Graph
StorageOpaque vector embeddingsFlat .md filesTyped graph, written as readable .md
RetrievalProbabilistic similarityDeterministic, whole-fileDeterministic traversal — node & edge
RelationshipsNone (lost in vectors)Implicit [[links]]Typed edges: depends-on, contradicts, causes…
RedundancyRe-embedded everywhereRepeats across filesDeduplicated & compressed (~88%)
Auditable / provenanceNoPartialYes — every edge cites its source
Accuracy vs RAG (F1)baselinebetter≈ 4× RAG (independently verified)
Chained reasoning (5-hop F1)0.170flat · no depth signal0.772 — rises with depth (v0.6.2)
Retrieval density (RDS)baseline42× vs RAG (v0.6.2)
Token costHighLower≈ 11× fewer than RAG
Failure modeSilently wrongLoudly missingLoudly missing — never silently wrong
Best for10M-doc archivesPersonal notes, one codebaseVertical enterprise reasoning, correctness-critical

F1/token figures from third-party benchmarking; RDS and 5-hop from the open benchmark v0.6.2. "GraphRAG" without compression benchmarks close to vanilla RAG — the gain is in the compression and typing, not the word "graph."

Two products people confuse — and shouldn't

Graphify · open-source tool

A thing a developer runs

Auto-extracts a graph from your own codebase and local docs. Developer-facing, code-centric, free DIY — you run it, tune it, own the upkeep. Great for giving a coding assistant memory of one repo.

Graphify.md · finished product

Intelligence you plug in

Delivers pre-built, verified vertical CKGs — finance, healthcare, manufacturing. No pipeline to run; drop the file in. Deterministic, auditable, benchmark-verified by construction. For a CIO who needs the answer right and provable, not just fast.

Plainly: a code tool turns your repo into a graph. Graphify.md turns an entire industry into reasoning-ready, verifiable intelligence. One is a utility; the other is the substrate enterprise AI reasons on.

When to use each — honestly

  • RAG — millions of documents, approximate recall acceptable.
  • LLM Wiki — personal notes or a single codebase you maintain by hand.
  • Compressed Knowledge Graph — when the answer must be right, dense, and defensible: regulated industries, high-stakes reasoning, anywhere a silently-wrong answer is a liability.
The hybrid that ships: CKG as the always-loaded, high-trust core; RAG as the fallback for the long-tail archive. Reason from the verified graph first; reach into the vector store only when the graph doesn't hold the answer.

Frequently asked

Isn't this just RAG with extra steps?

The opposite. RAG retrieves probably-relevant chunks from opaque vectors and can be silently wrong. A CKG traverses a readable, typed graph to the exact nodes and edges — deterministic, and when it doesn't have something it says so instead of hallucinating.

Isn't this just the Karpathy LLM Wiki?

The wiki is the folk version — a flat folder you hand-pile. A CKG is the engineered version: deduplicated, typed edges that carry the relationship, dependency-ordered, provenance-tagged, benchmarked. Same instinct, one rung up.

Isn't this just Graphify / the open-source graph tools?

Those are tools you run to graph your code. Graphify.md is a delivered product: finished, verified, vertical knowledge graphs for regulated enterprise reasoning — no pipeline to operate. A utility vs. an outcome.

Won't million-token context windows make this obsolete?

They make it better, not obsolete. Bigger windows let you load more of the right context — but dumping everything still costs tokens and buries the model in noise. Loading the precise, dense sub-graph wins on cost and accuracy regardless of window size.

Can I actually read and audit it?

Yes — it's Markdown. Every node and edge is human-readable and stamped with its source. Open it, diff it in Git, trace any claim to where it came from.

Stop piling files. Start reasoning on a graph.

See a Compressed Knowledge Graph next to your current RAG or wiki — and the benchmark behind it.

About Graphify.md

The problem was never the model. It's the space between.

A large model is brilliant at language and unreliable at truth — it has no verifiable memory to reason over, so it guesses, can't show its work, and can't tell you when it's guessing. The reflex has been to make the context window bigger. But the window was never the bottleneck: in a typical agent deployment, rules, orchestration, and RAG chunks consume ~85% of the budget before your company's actual knowledge enters the room. Graphify.md closes that gap — a compressed, owned, verifiable knowledge layer the model reads before it acts.

What we believe

Structure beats scale for truth

No amount of model scale produces grounding, provenance, or the right to refuse — those come from structure. A Compressed Knowledge Graph carries typed edges, provenance, and calibrated confidence, so an answer is traceable, not merely plausible.

How we work

We don't ship unverified claims

Every number on this site traces to an open, reproducible benchmark. A CKG can't assert what it can't support — refusal-with-evidence is a feature, not a failure. We hold our own output to the standard we sell.

Why us

Open standard, owned by you

The format, client, free CKGs, and benchmark are open — adopt them at no cost. The engine that turns your corpus into a faithful, current graph is what we sell. You own the layer; you rent the model.

Built by Dan Yarmoluk — inventor on three patent-pending filings (CKG compression, extraction, graph-grounded retrieval), author of the open benchmark, and maintainer of ckg-mcp, the production server distributing CKGs across dozens of domains. Graphify.md's context-budget thesis — the 25 / 30 / 30 / 15 breakdown — was recently featured by EnterpriseDB in CIO Dive.
The fastest way to see it on your data: a focused Domain Context Audit — we benchmark your domain (or a public one), show the token + accuracy delta vs your current RAG, and compile one pilot CKG you can run on any model. Start a pilot →

Baseline knowledge layer · token-optimized · model-agnostic

Instant corporate context. Your knowledge layer, in days.

A quick, model-agnostic knowledge layer — drop it into any model and your AI knows your domain the moment it answers. Your enterprise layer for better answers: agents 90% cheaper and 4× more accurate, and far stronger on chained reasoning — the linked, multi-fact questions where RAG breaks. Your knowledge already has a structure; we surface it, compress it, and you own it.

FastAccurateDynamicPersonalRelevantOwned
Scattered sources — structure already latent
DiscoverPressUse

The issue · context-window bloat

Your knowledge fights for ~15% of the window.

Every agent call re-ships the same overhead — rules, orchestration, retrieved chunks — before your domain knowledge even gets a seat. That's the bloat. A CKG compresses what matters into a few hundred tokens, so your knowledge isn't what gets squeezed out.

RAG · today
re-shipped chunks · 2,982 tok

The window fills with the same chunks every call — your knowledge gets squeezed.

CKG · with the layer
269
room to reason →

A dense, sourced graph — the window stays open for actual reasoning.

Rules 25%
Orchestration 30%
RAG chunks 30%
You 15%

~85% overhead before your knowledge enters the room (the 25 / 30 / 30 / 15 context budget). A CKG hands the model ~269 dense, sourced tokens instead — your knowledge, not the bloat. Cheaper to run, and the model reasons better with room to spare.

For the two who carry it

Need it now. Own it now.

Two people feel this before anyone else — the one who signs for the spend, and the one who gets paged when the agent is wrong. This is for both of you.

You sign for the spend · CFO

"AI cost is the one line I can't forecast — and it climbs every month."

It climbs because the model re-ships the same context on every call. Compress it once into a knowledge layer you own — instant corporate context, a fraction of the tokens, predictable spend, no lock-in. Your IP, on your books, enabled in days.

You get paged when it breaks · data / reliability

"Right at hour one. Confidently wrong by hour three."

That's drift and hallucination — a model guessing with no grounded memory. A CKG gives it one: 0% hallucination by construction, no context bloat, no drift. Deterministic traversal over a graph you start and manage — the inference you actually need. Sovereign: Claude, GPT, or a local model.

Need it, own it, right away — your context, your IP. An instant corporate-context layer you start and manage; sovereignty and agent success, by construction — without the bloat, the anxiety, the hallucination, or the drift.

Why now — and who feels it

Three forces, three needs, one layer.

Get the answer right first — then make it cheap, then make it yours. Accuracy, cost, ownership: for two years separate problems, now one layer that sits above the model. Each lands on a different desk.

Need 01 · Accuracy — knowledge expert & data engineer

Similarity isn't truth — and truth comes first

If the answer is wrong, cheap doesn't matter. RAG's documented failure is chained reasoning, relational reasoning — vector search finds similar text, not required text, and every hop compounds the error.

→ Traverse explicit edges: higher F1, accuracy that rises with complexity.

See it run →
Need 02 · Cost — CFO

Then it has to be cheap

AI bills for every query, and re-loading the same domain context on every call compounds until teams hit the wall by project five.

→ Compress it once: ~90% cheaper per answer — the context window stays free.

See the cost model →
Need 03 · Ownership — architect & CIO

And it has to be yours

As access tightens — retention, shifting policy, capability gating — routing every workload through a single external model concentrates risk prudent teams won't carry.

→ Own a portable layer: run it on Claude, GPT, or a local model.

See the architecture →
Each is a reason on its own. Together they're one architecture — a Compressed Knowledge Graph: compressed (cost), traversable (accuracy), and yours (ownership). Three pressures, one layer.
Watch all three resolve at once: a local 8B model goes from guessing to grounded — because the layer is compressed, correct, and owned. See it run →

Compress your context — own the knowledge layer. Rent the model.

Came from Karpathy’s LLM Wiki? See where the ladder leads — RAG → Wiki → CKG →

Open · documented · attested

A knowledge layer you can trust — and run yourself.

An owned layer earns trust three ways: the format is open, the build is documented, and the truth is signed by a human. Not a black box you rent.

Open standard  ✓ OKF-conformant

CKG Spec 1.0

The format is an open, ownable standard — typed nodes & edges, CSV / MD / TTL / JSON-LD. No lock-in: fork it, run it anywhere.

Read the standard →
Developer docs

Two lines to context

pip install ckg-mcp, four tools, the real schema, benchmark repro — live in Claude today.

Developer docs →
Ontology attestation

L2, signed by a human

A named domain authority reviews, prunes, and signs the graph — versioned, provenance-tracked, certified. Your name on the truth; the anti-black-box.

See attestation →

The issue · agents break down

Right at hour one. Confidently wrong by hour three.

It isn't a weak model — it's dark uncertainty: valid SQL, correct data, wrong answer, and no trail to debug. Agents drift as the session grows and re-guess on every hop. The fix isn't a bigger model; it's giving the agent a layer it traverses instead of a context pile it improvises over.

Agent on raw RAG

Drifts as context grows · re-guesses each hop · valid SQL, wrong answer · no audit trail — you're debugging a black box.

Agent on a CKG

Traverses declared, typed edges · cites every node · holds at depth · 0% hallucination by construction — and you can read the path it took.

A CKG is pre-action context — deterministic traversal, not generation. The agent stops guessing and starts citing. Same model, grounded. Watch it below — same 8B, same laptop, the only variable is the layer.

See it run — local, $0, no data egress

Same 8B model. Same laptop. The only variable is the knowledge layer.

We ran an open 8B model (qwen3:8b) on an Apple M4 / 16 GB — no internet, no cloud — and asked one hard, chained reasoning hardware-security question twice: once raw, once with a 9,377-token Compressed Knowledge Graph in context.

qwen3:8b — no CKG (raw 8B)
Q: How does a hardware attacker get
modified firmware to run past secure boot?

1. Secure Storage --[exfiltrated]--> Attacker
2. Attacker --[signs]--> Modified Firmware
3. Modified Firmware --[loaded]--> Bootloader
Tool: JTAG Debugger
Countermeasure: TPM + signing checks
✗ Invented structure · generic · not grounded
qwen3:8b + company CKG
Q: How does a hardware attacker get
modified firmware to run past secure boot?

FI1 --[defeats]--> BR14 --[enables]--> GA4
  fault-inject the signature check,
  then load the patched image
Lab tool:        LAB4 (ChipWhisperer)
Countermeasure:  HHC7 (Protect/Detect/Respond)
Every hop is a sourced edge.
✓ Grounded · traceable node IDs · 0% invented
Run it yourself — any laptop, ~10 min, $0
brew install ollama          # one time (or ollama.com)
ollama pull qwen3:8b         # ~5 GB · runs on an M-series Mac / 16 GB
python3 demo.py              # same question — with vs. without the CKG

No GPU, no cloud, no API key — nothing leaves your machine. demo.py + the CKGs ship in the bundle and the benchmark repo. Prefer zero setup? Paste a free CKG into any model and ask a hard question — or run the full open benchmark across 45 domains. If you can install one app, you can reproduce this.

Raw, the model improvised — generic, ungrounded, impossible to trace to a source. With the compressed knowledge layer it answered along the graph's explicit edges and cited every node, so you can check each one against the file. Fully local: nothing left the machine, cost ≈ $0. That is "own the knowledge layer, rent the model" — running on a laptop.

Architecture · how it works

A layer above the model — not more context inside it.

The context window is finite and metered, so the answer isn't a bigger window — it's denser knowledge, put where it belongs: above any model, owned by you, and delivered before the agent acts.

Your corpus
docs · standards · code · data
↓ engine: discover · compress · verify (built once)
Compressed Knowledge Graph
open format · you own it
↓ pre-action context · ~269 tokens · traversed, not searched
Any model
frontier · open · local — swap freely
↓ grounded · cited · bounded
Agents & apps
deterministic answers · 0% hallucination

One question, two ways.

"Does your service work mid-onboarding with compliance requirements?"

RAG — similarity search

Finds chunks about onboarding + compliance + services, then synthesizes a link that was never stated. The model invents the relationship.

precision ~60% · recall ~25% · hallucination risk high
CKG — graph traversal

Walks explicit, typed edges. The answer cites its path — every hop is a sourced edge.

Service →[applies-to]→ ServiceTier →[supports]→ MidStage →[needs]→ ComplianceCheck ✓
OPEN MCP · 01

query_ckg

Typed nodes + edges, not prose.

OPEN MCP · 02

get_prerequisites

Dependency chains, resolved.

OPEN MCP · 03

traverse

Walk N hops from any node.

OPEN MCP · 04

validate_ckg

Check against a profile (see Certification).

Where RAG and ontologies fit: RAG retrieves passages, an ontology validates meaning — the CKG is the reasoning layer between them, and the only one you fully own. Knowledge arrives before the agent acts — pre-action context, not mid-task retrieval.

See the discipline — interactive

Watch the architecture form, then compress.

Drag the meter: raw corpus → discovered ontology → typed dependencies → a CKG that compresses to a fraction of the tokens as the context window literally opens. Same graph, fidelity intact.

Open proof, not claims

Open benchmark: 7,928 queries, 45 domains, reproducible.

Macro-F1 across Track 1's 44 domains (45 with the GLP-1 transfer track), fully reproducible. Code, data, and paper public — run it yourself before you talk to us.

What good looks like — the rubric

Don't grade on one number. Judge the whole ladder.

RAG → GraphRAG → DAG → CKG. Correctness, cost, energy, depth, and trust — together. (CKG / RAG / GraphRAG from benchmark v0.6.2; DAG = the structural rung a CKG perfects.)

Metric"Good" =RAGGraphRAGCKG
Accuracy — Macro-F1right, not just fluent0.1230.1200.471
Tokens / querydense, minimal2,9823,450269
Cost / querycents, predictable~$0.0106~$0.0166~$0.0010
Cost per finished taskbounded, one passcompounds/hopcompounds/hopone traversal
Retrieval Density Score (RDS = F1 ÷ tokens)high signal / token<1×~42×
Intelligence per wattmore answer / joulebaselineworse (+tokens)~11× fewer tokens†
Inferencefast, deterministicembed + top-k+ community searchdeterministic traversal
Depth (multi-hop F1)holds / rises with hopsflat ~0.07–0.230.37 → 0.77
Semantics — BERTScorefluent ≠ correct0.8170.857
Hallucinationzero by constructionunboundedinvents edges0%
Traceabilitynode-level citationpartial
Reproducibilityclone & re-run✓ open
Ownershipportable, localpipelinepipeline
Read the accuracy row right — it's the thesis, not a contradiction. A derived graph (GraphRAG, 0.120) scores about the same as plain RAG (0.123): bolting on "a graph" doesn't move accuracy. The 4× jump is CKG (0.471) — structure that's authored, compressed, and traversable, not extracted on the fly. DAG is the ideal sitting between them; a CKG is that ideal realized. It isn't the word "graph" that wins — it's the graph done right.

†Intelligence-per-watt is directional (fewer tokens → proportionally less inference energy); a metered watts-per-answer figure is in measurement — we won't quote one we haven't run. RDS isn't a third trophy — it's F1 × token-efficiency on one axis (≈ 3.8 × 11). BERTScore near-parity is the point: both sound right; Macro-F1 is where "is right" diverges. Figures v0.6.2 (Macro-F1 = Track 1, 44 domains); cost & 5-hop per-query. Re-run: github.com/Yarmoluk/ckg-benchmark.

Reading the ladder — and the "GraphRAG is cheaper" myth

RAG → GraphRAG → DAG → CKG: what each actually is.

RAG

Retrieve text chunks by embedding similarity. No structure — the model re-reasons over a pile of prose every call.

GraphRAG

Dynamically extracts a graph from text at index time (LLM entity extraction + community detection), then retrieves over it. Real and shipped — but the structure is derived, not known.

DAG (the ideal, not a product)

The architectural shape good structured knowledge takes: a directed acyclic graph of typed concepts. A target, not a vendor.

CKG

The DAG done right: pre-structured (declared, not extracted), compressed, and attested. No derivation step to pay for.

"But isn't GraphRAG cheaper long-term?" The claim assumes you build the graph once, then queries get cheap. Our data shows the opposite — GraphRAG pays twice: an expensive build (LLM extraction + community detection over the whole corpus) and expensive queries (~3,450 tokens each — more than plain RAG, because it routes local-vs-global community search). The amortization never arrives. A CKG inverts both: ~zero build (structure is authored, not derived) and ~269 tokens/query. That's why the table shows GraphRAG as not cheaper — clone the benchmark and confirm it.

The ruler, not the promise. Like NVIDIA rates silicon by performance-per-watt — a ruler, not a guarantee — judge a knowledge layer by intelligence per token (RDS) and per watt: efficiency you can measure, on your own corpus. The rubric is the ruler; your data is the verdict.

MetricCKGRAGGraphRAG
Macro-F10.4710.1230.120
Tokens per query2692,9823,450
Run cost · full benchmark$7.81$76.23$44.43*
F1 at 5-hop complexity0.7720.170
Hallucination rate0% by constructionunmeasuredunmeasured

GraphRAG burns more tokens than the RAG it claims to replace (3,450 vs 2,982) — traversal overhead eats the gains. *Its lower run total is a 15-domain subset, not apples-to-apples. Retrieval Density Score: ~42× over RAG.

The ladder: RAG → GraphRAG → DAG → CKG. RAG finds similar text, probabilistically. GraphRAG is the graph layer most enterprises reach for — but it builds a graph then still summarizes & retrieves probabilistically (LLM community reports), so it ties plain RAG (0.120 vs 0.123) and burns more tokens (3,450 vs 2,982). A DAG — a typed, acyclic graph you traverse deterministically — is the structure that actually beats GraphRAG. A Compressed Knowledge Graph is that DAG, done right: compressed, provenance-tagged, ~269 tokens. The benchmark measures CKG vs RAG vs GraphRAG — CKG wins every column. The DAG is the why; the CKG is the DAG you can afford and audit.

The Structure Paradox

More structure should mean more rigidity. Instead, accuracy rises with complexity.

Answer accuracy (F1) 00.51.0 question complexity (hops) CKG 0.772 RAG 0.170 1235

Answers become overdetermined, not uncertain. Every hop makes RAG guess again; every hop makes CKG more certain.

The harder the question, the wider the gap

By question type (CKG vs RAG)CKGRAG
Dependency0.6340.078
Chained-reasoning path0.6600.201
Category aggregation0.9640.286

Also scored with semantic BERTScore F1 (roberta-large): 0.857 vs 0.817 (RAG) — a tighter spread because the metric is lenient. The exact Macro-F1 above is the harder, more honest number.

Energy & inference

Inference latency

<100 ms

p95, live POC. Graph traversal is cheap deterministic compute — not another generation pass.

Intelligence per watt

~11× fewer tokens

= proportionally less inference energy per correct answer. A metered watts-per-answer figure is in measurement — we won't quote one we haven't run.

Local-model proof

8B · on a laptop

An 8B model + a CKG out-reasons a far larger model on chained reasoning — fully offline, ≈ $0, no data egress.

What we're still measuring

A metered intelligence-per-watt figure · GraphRAG on the full 45-domain set (current is a 15-domain subset) · BERTScore reconciliation to v0.6.2. We publish what's measured and flag what isn't — that discipline is the product.

Three ways in · one knowledge layer

Same layer. Three doors — because you're not all here for the same reason.

A marketer tired of being invisible to AI. A developer who's watched RAG drift. A platform owner who needs governance, not a demo. The CKG underneath is identical — walk through the door that's yours.

SEO & Marketing · GAO

"AI answers questions about our category — and never names us."

Be the source the models cite, not the page nobody clicks. Publish your knowledge as a CKG surface AI retrieves first. White-label for agencies.

Make it AI-discoverable →
Developers & Agents · MCP

"I've shipped RAG. I've watched it drift by hour three."

Context that just shows up — no pipeline, no embeddings to tune. pip install ckg-mcp and the tools appear inside Claude (Desktop & Code) today.

Connect an agent →
Enterprise & Platform · API

"I can't put a science project into production."

Keys, tiers, metering, SLAs, audit logs — the same engine behind a governed HTTP surface you can take through security review.

Get an API key →

Developer docs

Two lines to first context.

MCP is live today (pip install ckg-mcp). The REST surface is provisioned per engagement — same core, same method, behind a key.

MCP · Claude (Desktop & Code)live · 4 tools
# pip install ckg-mcp
{ "mcpServers": { "ckg": {
    "command": "ckg-mcp" } } }
# query_ckg · get_prerequisites
# search_concepts · list_domains
REST · POST /v1/queryper engagement
curl -X POST .../v1/query \
  -H "Authorization: Bearer $CKG_KEY" \
  -d '{"domain":"glp-1",
      "question":"renal contraindications?"}'
# → answer + cited path · tokens: 269

Developer docs

Two lines to first context.

MCP is live today (pip install ckg-mcp). A CKG ships as typed CSV / Markdown — query it from Claude (Desktop & Code), or over REST per engagement. Same core, same method, behind a key.

1 · Install & connect

MCP · Claude (Desktop & Code)
# 1 · install
pip install ckg-mcp

# 2 · add to your MCP config
{ "mcpServers": { "ckg": { "command": "ckg-mcp" } } }

# 3 · the CKG tools now appear inside your agent

2 · The four tools

list_domains

What knowledge is available — the bundled catalog of domains.

query_ckg(domain, concept, depth)

Traverse the graph — returns typed nodes + edges, not prose.

get_prerequisites

What must be understood first — dependency chains, resolved.

search_concepts

Find the right entry point across the ontology.

3 · REST (provisioned per engagement)

request · POST /v1/query
curl -X POST .../v1/query \
  -H "Authorization: Bearer $CKG_KEY" \
  -d '{ "domain":"glp-1",
        "question":"renal contraindications?" }'
response · ~84 ms
{ "answer": "constrained below eGFR…",
  "path": ["Semaglutide→RenalDosing→…"],
  "edges_cited": 3,
  "tokens": 269,
  "hallucination": "0 — traversal, not generation" }

4 · Formats & conformance

Every CKG exports as .md, .ttl (RDF/Turtle), or JSON-LD — readable, Git-versionable, model-agnostic. validate_ckg checks output against a conformance profile: pass = L1 (structurally valid); a domain authority reviews & signs = L2 (certified). The discovery/compression method stays server-side — you get results and format, never the recipe.

5 · Power users — the ontologist's loop (approve, don't extract)

Your job is approval, not extraction — the engine handles discovery. Author a conformance profile (the rules your domain must obey), then review the engine's output in the console: validate_ckg confirms structural conformance (L1); you approve · edit · prune · add — every node carries confidence + provenance — then sign for L2, named and versioned (e.g. pursuit-ckg@1.2). The anti-black-box: nothing ships that you didn't see.

6 · The CKG format — the open standard

A CKG is a directed acyclic graph (DAG) of typed concepts with explicit prerequisite edges — declared, not inferred. No graph database, no embeddings. It serializes as plain-text CSV or Markdown (exportable to RDF/Turtle and JSON-LD): readable, Git-diffable, model-agnostic.

learning-graph.csv · node schema
ConceptID,ConceptLabel,Dependencies,TaxonomyID
1,Function,,FOUND
2,Domain and Range,1,FOUND
3,Limit,2,CALC
# Dependencies = pipe-delimited prerequisite ConceptIDs → the edges.
# Acyclic by construction. TaxonomyID groups concepts.

Query types the standard supports: T1 entity lookup · T2 direct dependency · T3 multi-hop path · T4 category aggregate · T5 cross-concept relationship. Conformance: L0 raw → L1 structurally valid (validate_ckg against a profile) → L2 authority-signed.

7 · Reproduce the benchmark

clone & run · MIT code · CC BY 4.0 data
git clone https://github.com/Yarmoluk/ckg-benchmark
cd ckg-benchmark
pip install -r evaluation/requirements.txt
python evaluation/ckg_harness.py --domain calculus
python evaluation/rag_harness.py  --domain calculus
python evaluation/analyze_results.py

Scope: 45 domains · 7,928 queries (Track 1: 44 educational, 7,758 q; Track 2: GLP-1, 170 q) · Macro-F1 0.471 / 0.123 / 0.120 (Track 1) · RDS ~42× (mean of per-domain F1÷tokens) · pre-print in preparation.

Context Architecture · company-owned AI infrastructure

Your context architecture — owned, not rented.

One context layer above every model and agent — open, governed, portable, and a Compressed Knowledge Graph underneath. Adopt it through whichever door fits your team; the infrastructure is yours.

CFO · Finance

Token Economics

Stop re-billing the same context every call. Compress domain knowledge once; up to ~42× more token-efficient per query.

See the cost model →
Developer · Engineering

Agent Reliability

Your agent is right at hour one, confidently wrong by hour three — and you can't debug a black box. Give it a CKG as pre-action context: deterministic traversal, every answer cited. No drift, no context rot, 0% hallucination by construction.

See it run on a local 8B →
Domain Expert · Ontologist

Context Conformance

Don't rebuild graphs by hand — we run the discovery engine for you. You review, prune, and certify the output in a console: you sign, it ships L2. Your name on the truth, none of the grunt work.

See the review console →
Marketing · Content

GEO — Generative Engine Optimization

Be the source LLMs cite, not the page they skim. Publish your knowledge as a CKG surface AI retrieves first. (patent-pending)

Make it AI-discoverable →
Enterprise Architecture

Semantic Substrate

One owned, model-agnostic layer above every model and agent — open format, governed, auditable, no single point of failure.

See the architecture →
Data & Knowledge Teams

Automated Discovery — the Engine

Point it at your corpus; get a faithful, traceable CKG in a session, not eight days. .md / .ttl / JSON-LD.

Run it on your corpus →
Same open standard underneath, six products on top — adopt one door, gain the rest. Book a pilot →

Open by design — CKG Spec 1.0 · ✓ OKF-conformant

Open the standard. Keep your knowledge. Build what you can't.

Graphify.md is open-core, with a hard line down the middle: the format is an open, ownable standard; the hard part — turning a corpus into a faithful, current graph — is the patent-pending engine we run for you.

OKF-conformant. Google's Open Knowledge Format (June 2026) standardized exactly the container a CKG lives in — knowledge as portable markdown. Every CKG is a valid OKF bundle; not every OKF bundle is a CKG. OKF gives you portability; the CKG layer adds what OKF deliberately leaves out — compression, typed relationships, deterministic traversal, and L2 attestation. The same files Google can ingest — semantic-layer-grade inside.

See it — OKF in, CKG out

Plain OKF document (Google's spec)
---
type: Concept
title: GLP-1 Prior Authorization
description: Payer gate before a GLP-1 is dispensed.
tags: [glp1, payer, access]
timestamp: 2026-06-18T00:00:00Z
---
# Prior Authorization
Most plans require step therapy and a documented
BMI / comorbidity before approving a GLP-1.
The same file as a CKG — still valid OKF
---
type: CKG Concept       # the one OKF-required field
title: GLP-1 Prior Authorization
tags: [glp1, payer, access]
timestamp: 2026-06-18T00:00:00Z
ckg:                  # extension — OKF readers ignore
  id: 42
  confidence: 0.94
  source: ClinicalTrials.gov + 2026 policy
---
# Prior Authorization — typed edges, not prose
depends_on: Step Therapy(38) | BMI≥30(31)
enables:    Dispense GLP-1(45)
gated_by:   Formulary Tier(40)

Left: a Google OKF doc — portable, but prose. Right: a CKG — same frontmatter (a plain OKF reader, including Google's Knowledge Catalog, ingests it unchanged), plus typed edges, confidence, and provenance. Portable on the outside; deterministic + sourced inside. Download this sample →

OKF documentation & announcement. Google Cloud announced the Open Knowledge Format on June 12, 2026 (Sam McVeety & Amir Hormati, Data Cloud) — the announcement · the spec & sample bundles. Our documentation — a compressed knowledge graph that is itself a valid OKF bundle: OKF × CKG (.md) ↓.

Part one — the open standard

Yours to own, fork, run anywhere.

  • CKG format / spec — typed nodes & edges, .md+.ttl+JSON-LD.
  • Reference client & validatorckg-mcp.
  • The benchmark — code, data, paper.
  • Free starter CKGs — no email gate.
MIT · specApache-2.0 · clientCC BY · graphs

Part two — the engine

The part you can't build yourself.

  • Automated ontology discovery + compression — corpus → CKG in one session.
  • Fidelity & grounding — every node traceable to source.
  • Premium domain CKGs — curated, cross-document.
  • Enterprise serving — private, hosted, fresh, governed.
Patent-pending engineCommercial
Open everything that helps you own and use a CKG. We earn our keep on the part that produces one — faithfully, fast, and current. A hand-built graph took one team eight days for a single book; the engine does it in a session.
What's live today: the benchmark paper, the open ckg-mcp client, free CKGs, and 54 bundled domains. CKG Spec 1.0 is in final review ahead of public release — the paper is the canonical reference until it ships.

Trust on demand — the authority in the loop

Auto-discovered, expert-certified.

Automated discovery gives you a scaffold; a semantic authority makes it trustworthy. Graphify.md turns that into a configuration surface — so the ontologist becomes a certifier, not a from-scratch builder. Assistance, not replacement — made mechanical.

01 — CONFIGURE

Conformance Profile

The authority authors the rules once: vocabulary, SHACL shapes, ontology alignment, domain axioms.

02 — VALIDATE

Conformance check

The engine's CKG is validated against the profile → a SHACL-style report (conforms: true). Only findings need review.

03 — CERTIFY

Signed certification

The authority certifies it. The signed certification travels inside the CKG — verifiable offline by any consumer or auditor.

L0

Core

Valid format + provenance. What the engine emits.

L1

Conformant

Machine-validated against a Conformance Profile (SHACL conforms).

L2

Certified

Signed by a semantic authority / data steward. The tier enterprises buy.

Profile, validation, and certification formats are open (verify trust without us); the discovery engine stays proprietary. "Certified by a semantic authority" is a sellable trust mark — the same pattern data stewards use to certify trusted assets in Collibra, Alation, and Power BI.

The discipline

Automated ontology discovery + compression.

Not prompt-stuffing. Not chunk-and-pray. A repeatable method: discover the ontology latent in a corpus, map explicit typed dependencies across documents, and compress to a fraction of the tokens with fidelity intact.

01 — DISCOVER

Surface the ontology

The structure documents imply but never state.

02 — CONNECT

Typed dependencies

Explicit, sourced edges — including cross-document links RAG can't see.

03 — COMPRESS

Distill with fidelity

~269 tokens/query of bounded, traceable context.

04 — VERIFY

Ground to source

Every node carries provenance. 0% hallucination.

The proof of rigor underneath: three peer-style intelligent textbooks (Context Engineering · Agent Memory Patterns · Ontology Engineering — 540+ concepts) and an open, reproducible benchmark. By Daniel Yarmoluk with Dan McCreary (creator of intelligent textbooks). The engine is patent-pending; the format it produces is open.

Free CKG Library — 53 ship in ckg-mcp · 45 benchmarked · no gate

Open your window. Dozens of free knowledge graphs.

Every graph here is free — render it in the viewer, grab one as a .md (no signup), or pip install ckg-mcp for all 53 (45 formally benchmarked, v0.6.2 — clone & re-run; the rest staged). Paste it into Claude, ChatGPT, Gemini, or a local model, or let it appear as an MCP tool. (What is paid is the engine that turns your corpus into the next one.)

Render & download any CKG — free (one per person), or pip install for all:

Drag to explore · scroll to zoom · pick from the dropdown, or click any card below.

These are public, education- and reference-domain CKGs that prove breadth. Your private corpus becomes an L2 certified CKG via the engine — same format, never leaves your control.
Need a CKG of your company or your website? That's the engine — point it at your corpus and get a faithful, traceable CKG you own. Contact us →

Part 2 · Get started

Get your knowledge layer running in days.*

Start free with the open standard and a CKG; bring the engine in when you want your own corpus turned into a faithful, current, certified graph. Own the standard. Own your knowledge. Rent the model.

* A scoped, single-domain CKG — running in days. Auto-discovery does the heavy lifting in a session; you review and certify (L2); production hardening follows. Not a year-long platform project.

Where to start · how to evaluate · live in days

Start narrow. Evaluate hard. Ship in days.

1 · Where to start

One pressing domain — the proposals, the support corpus, the spec library that hurts most. We map it free (a Domain Context Audit) and show the delta on your data, not a slide.

2 · How to evaluate

Don't take our word — run the bake-off. On your hardest multi-hop questions, compare LLM-only / RAG / GraphRAG / CKG on accuracy (F1, groundedness, citations), tokens/query, cost per correct answer, and whether it holds at 5 hops. We hand you the harness; demand the same of any vendor.

3 · Live in days

Auto-discovery builds the graph in a session; you review & certify; it drops into any model — pip install ckg-mcp, plug and go. Days, not years. Way fewer tokens, way lower cost, owned by you.

Free · Open

The standard

Read the paper, run the benchmark, grab a free CKG, pip install ckg-mcp. (CKG Spec 1.0 in final review.)

Go to GitHub →
Build · API

The engine

Turn your corpus into a CKG via API — `.md` / `.ttl` / JSON-LD, on any model.

Talk to us →
Enterprise · Certified

L2 graphs

Private, hosted, governed, and certified by a semantic authority. No data egress.

Book a pilot →

Graphify.md · Compressed Knowledge Graphs · Patent-pending engine (US utility applications filed) · hello@graphifymd.com