Your coding assistant writes LangChain code from 2024. LangChain v1 renamed create_react_agent → create_agent, moved it to langchain.agents, renamed prompt → system_prompt, and exiled legacy chains to langchain-classic — and every model trained before late 2025 emits the old API by default, while vector indexes of old tutorials confirm it.

This package ships the antidote inside the wheel: pre-built knowledge graphs of the agent stack — LangGraph (including the v1 migration as typed REPLACES edges), MCP, agent memory, the major agent platforms — where every node carries the SHA-256 of the docs page it was extracted from. "Is this current?" becomes a mechanical check, not a hope.

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

Installation

pip install langchain-ckg

Depends only on langchain-core and httpx. The 11 graphs add ~120KB to the wheel.

Instantiation

from langchain_ckg import CKGRetriever, available_domains

retriever = CKGRetriever()          # bundled langgraph domain, offline
print(available_domains())          # all 11 bundled domains

Usage

docs = retriever.invoke("checkpointer")
print(docs[0].page_content)
# # Checkpointer (langgraph)
# ## Prerequisites
#   - MemorySaver
#     - StateGraph
#       - LangGraph Framework
# ## Builds toward
#   - thread_id
# Source: https://docs.langchain.com/oss/python/langgraph/persistence
# Source hash: sha256:114c96d9...

print(docs[0].metadata["source_url"])   # provenance in metadata too

Verify any answer against the live docs — no trust required:

curl -s https://docs.langchain.com/oss/python/langgraph/persistence | shasum -a 256
# equals the stored source_hash, or the edge is stale — deterministic either way

Use within an agent

from langchain.agents import create_agent
from langchain.tools import tool
from langchain_ckg import CKGRetriever

retriever = CKGRetriever()

@tool
def langgraph_map(query: str) -> str:
    """Look up current LangGraph concepts, their prerequisites, and source docs."""
    return "\n\n".join(d.page_content for d in retriever.invoke(query))

agent = create_agent(model="anthropic:claude-opus-5", tools=[langgraph_map])

One-liner alternative: from langchain_core.tools import create_retriever_tool (works on langchain 0.3.x and 1.x).

Bundled domains

Examples: examples/stale_api_correction.py · examples/agent_memory_map.py — both run offline with no API key.

Hosted retriever — the full 100+ domain library

CKGHostedRetriever queries a hosted CKG MCP endpoint (NVIDIA stack, finance, healthcare, compliance domains). 48-hour free window, then 402 with upgrade options.

from langchain_ckg import CKGHostedRetriever

retriever = CKGHostedRetriever(domain="nvidia-nemo")            # free 48h
retriever = CKGHostedRetriever(domain="nvidia-nemo",
                               license_key="CKGAP-...")         # graphifymd.com/pricing

# x402 — agent pays itself in USDC on Base L2 when it hits a 402
retriever = CKGHostedRetriever(
    domain="nvidia-nemo",
    x402_private_key="0x<evm-private-key>",   # pip install 'langchain-ckg[x402]'
)

# Lightning — pre-paid invoice, ~$0.001/call
retriever = CKGHostedRetriever(
    domain="nvidia-nemo",
    lightning_invoice_id="<invoice-id>",      # GET {endpoint}/lightning/invoice
)

Trust anchor chain

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

source_url:  https://docs.langchain.com/oss/python/langgraph/...   ← fetch hint
source_hash: sha256:<64-char hex>                                  ← trust anchor

Full audit chain: edge answer → graph commit → source_hash → source_url. A hash mismatch means the upstream docs changed — the graph tells you it's stale instead of quietly guessing. Four bundled domains (agent-memory, agent-loop-patterns, crewai, llamaindex) currently carry extraction-internal references (metadata.provenance = "extraction-internal") pending re-anchor to public URLs; the rest verify with curl today.

API reference

CKGRetriever(domain="langgraph", depth=3, k=5) — bundled/local, sync .invoke() (async via default ainvoke delegation). CKGHostedRetriever(endpoint=..., domain=..., license_key=..., x402_private_key=..., lightning_invoice_id=..., depth=3, k=5) — hosted MCP. available_domains() -> list[str] — bundled domain names.

Links

graphifymd.com · Pricing · PyPI · Benchmark · Dataset