CKG vs RAG — the difference that matters

RAG                                    CKG
────────────────────────────────────   ─────────────────────────────────────
Retrieve document chunks               Traverse typed dependency graph
Probabilistic similarity match         Deterministic BFS from matched concept
No provenance                          SHA-256 per node — verify any claim
2,982 tokens/query                     269 tokens/query  (11× cheaper)
F1 = 0.123                             F1 = 0.471  (4× better)

Benchmark paper · Dataset · patent-pending methodology

Install

pip install langchain-ckg

Usage

Local retriever — free, no network

Queries run entirely in-process against graphs bundled in ckg-mcp. No API key, no rate limit.

from langchain_ckg import CKGRetriever
from langchain_openai import ChatOpenAI
from langchain.chains import RetrievalQA

retriever = CKGRetriever(domain="nvidia-nemoclaw", depth=3)
qa = RetrievalQA.from_chain_type(llm=ChatOpenAI(), retriever=retriever)
result = qa.invoke("How does NemoClaw handle CUDA kernel fusion?")

Hosted retriever — 48-hour free trial, then $1 / 100 calls

from langchain_ckg import CKGHostedRetriever

# Free for 48 hours from first call
retriever = CKGHostedRetriever(domain="nvidia-nemoclaw")

# After trial: pass a license key
retriever = CKGHostedRetriever(
    domain="nvidia-nemoclaw",
    license_key="CKGAP-...",  # graphifymd.com/pricing
)

After 48 hours the server returns 402 Payment Required with upgrade options. See Payment rails for autonomous payment.

Payment rails

x402 — autonomous USDC on Base L2

Agent pays itself. No human in the loop.

retriever = CKGHostedRetriever(
    domain="nvidia-nemoclaw",
    x402_private_key="0x<evm-private-key>",
)
# On 402: signs $0.010 USDC payment, retries, returns result

License key — human-purchased, unlimited

graphifymd.com/pricing → receive CKGAP-... key → pass once:

retriever = CKGHostedRetriever(domain="nvidia-nemoclaw", license_key="CKGAP-...")

Lightning — autonomous sats

retriever = CKGHostedRetriever(
    domain="nvidia-nemoclaw",
    lightning_invoice_id="<invoice-id>",  # GET {endpoint}/lightning/invoice
)
# 100 sats/call (~$0.001)

Metered billing with PolarUsageCallback

Track per-retrieval usage in Polar — fires a background thread, never blocks the chain:

from langchain_ckg import CKGRetriever, PolarUsageCallback

cb = PolarUsageCallback(api_key="sk_...", external_customer_id="license_key_abc")
retriever = CKGRetriever(domain="nvidia-nemoclaw", depth=3)
qa = RetrievalQA.from_chain_type(llm=ChatOpenAI(), retriever=retriever, callbacks=[cb])

AgentKit integration

Coinbase AgentKit agents pay x402 natively. Wire CKGHostedRetriever as a LangChain tool:

from coinbase_agentkit_langchain import CoinbaseToolkit
from langchain_ckg import CKGHostedRetriever
from langchain.tools.retriever import create_retriever_tool

retriever = CKGHostedRetriever(
    domain="nvidia-nemoclaw",
    x402_private_key="0x<agentkit-wallet-key>",
)
ckg_tool = create_retriever_tool(
    retriever,
    name="query_knowledge_graph",
    description="Retrieve SHA-256 anchored knowledge about NVIDIA NemoClaw. Pays autonomously via x402.",
)
# Combine with AgentKit tools and run

Trust anchor chain

Every node carries a SHA-256 hash of its source document bytes at extraction time:

source_url:  https://docs.nvidia.com/nemo/...   ← fetch hint
source_hash: sha256:<64-char hex>               ← trust anchor

Verify any claim:

curl -s <source_url> | sha256sum
# mismatch = stale edge or upstream silent edit — no judgment needed

Full audit chain: edge answer → graph commit → source_hash → source_url

Available domains

117 domains — NVIDIA AI stack, NemoClaw, Salesforce AgentForce, finance (Basel III · SEC · IFRS), agent protocols (MCP · A2A · x402), infrastructure (Render · Stripe · PostHog · Cloudflare), and more.

uvx ckg-mcp  # → call list_domains

Links

graphifymd.com · Pricing · PyPI · Benchmark · Dataset