ckg-nvidia-ai · Graphify.md

You've hit the limit. Here's what you gained.

Same question. Different answer quality.

Q: "What does NIM depend on to handle model routing?"

RAG → NVIDIA NIM is a set of cloud-native microservices that simplifies
       deployment... NIM leverages TensorRT and Triton... routing involves
       profiles selected at runtime... [2,847 more tokens, source unknown]

CKG → { "concept": "nim-model-routing",
         "depends_on": ["triton-inference-server", "nim-profile-selection"],
         "source_hash": "sha256:a3f8c2...",
         "tokens": 247, "tokens_saved": 2735 }

Structured, sourced, 11× cheaper. Every time.

Before — RAG

Tokens / query2,982

Accuracy (F1)0.123

Cost / 1k calls$8.95

Answer sourceInferred

After — CKG

Tokens / query269

Accuracy (F1)0.471

Cost / 1k calls$0.81

Answer sourceSHA-256 verified

The math

Token cost per 1,000 queries (at $0.003 / 1k tokens)
RAG    2,982 tokens  ×  $0.003/1k  ×  1,000  =  $8.95
CKG      269 tokens  ×  $0.003/1k  ×  1,000  =  $0.81
                                     savings  =  $8.14 per 1k queries  (11×)

Accuracy — F1 on NVIDIA AI domain queries
RAG  0.123  ██░░░░░░░░░░░░░░░░░░░░  wrong 87.7% of the time
CKG  0.471  ████████░░░░░░░░░░░░░░  wrong 52.9% of the time
     ckg-benchmark v0.6.2 · 28 queries · huggingface.co/datasets/danyarm/ckg-benchmark

Your trial — 150 calls

—

tokens saved

—

RAG cost avoided

4.7×

fewer wrong answers

Your annual projection

$10 / year · $0.83/month

Get your license key — arrives in ~10 seconds

Pay once. Use all year. Not useful within 30 days? Reply to the key email for a refund, no questions.

Haven't tried it yet? Start free → How to connect →

After you pay — 3 steps 1. Key arrives in your inbox in ~10 seconds.
2. Add X-License-Key: CKG-xxxx-xxxx to your MCP client config.
3. No call limit. All 20 domains. Every query still shows tokens_saved.

Works with AG2 AutoGen crewAI LangChain LlamaIndex Claude Desktop Cursor Hugging Face

▶ New to MCP? Connect in 2 minutes

MCP is how AI agents connect to external tools. Add this to Claude Desktop, Cursor, or Windsurf:

Option A — URL (claude.ai Connectors)

https://ckg-nvidia-ai.onrender.com/mcp

Option B — JSON config (Claude Desktop / Cursor)

{
  "mcpServers": {
    "ckg-nvidia-ai": {
      "command": "uvx",
      "args": ["ckg-nvidia-ai"]
    }
  }
}

Then ask: "List the available NVIDIA AI domains" or "What does NIM depend on?"

Daniel Yarmoluk, Graphify.md
Built this because agents were burning tokens on knowledge they should already have structured. Reply to the key email — I read every one.
README + changelog →