ckg-nvidia-ai · Graphify.md

What it is and how to use it

A knowledge graph of the NVIDIA AI stack, served as an MCP tool your agent calls directly. Instead of embedding text and searching vectors, your agent traverses declared edges between concepts. Faster, cheaper, same answer every time.

269 mean tokens vs 2,982 for RAG 4.7× better F1 20 domains · 998 nodes

Step 1 — Connect

Add to your MCP client (Claude Desktop, Cursor, or any MCP-compatible agent):

{
  "mcpServers": {
    "ckg-nvidia-ai": {
      "url": "https://ckg-nvidia-ai.onrender.com/mcp"
    }
  }
}

Or run it locally:

pip install ckg-nvidia-ai
ckg-nvidia-ai

Step 2 — Use the tools

list_domains

See all 20 NVIDIA AI domains available (NIM, NeMo, CUDA, TensorRT, Triton, Guardrails…)

search_concepts "quantization"

Find concepts by keyword within a domain

query_ckg "tensorrt-llm-quantization"

Traverse the graph outward from a concept — returns edges, dependencies, source hashes

get_prerequisites "nemo-curator"

Full upstream dependency chain — everything your agent needs to know before acting on this concept

ask_nvidia "how does NIM handle routing?"

Natural language → graph-grounded answer, traced to source URLs

Step 3 — Verify a source (optional)

Every node carries a SHA-256 hash of its source page. Recompute at any time:

curl -s https://docs.nvidia.com/nemo/... | sha256sum
# compare to source_hash in the response

Mismatch = upstream page changed. The graph tells you which edges to re-extract.

Free tier

50 calls/day. Upgrade for unlimited →