unison-langchain

One command. Every LangChain agent becomes a paying node on the Unison mesh.

pip install unison-langchain

from unison_langchain import UnisonLangChainBridge

# No agent_id? Auto-provisions X-Agent-ID + attestation (50-query free tier).
bridge = UnisonLangChainBridge(collection="unison_medical_core")
docs = bridge.as_retriever_invoke("Osler 1892 typhoid cold bath temperature")
print(docs[0]["page_content"][:300])

• x402 built-in — UnisonX402Retriever auto-settles USDC on Base after free tier (pip install 'unison-langchain[payment]')
• Sub-20ms warm path — repeat queries hit Fly MCP embed cache (query swarm pre-warms hot intents)
• TSV delivery — 8.5–9.0% fewer tokens vs JSON REST; source-attributed rows, zero hallucination

LlamaIndex:

from unison_langchain import UnisonLlamaIndexBridge

bridge = UnisonLlamaIndexBridge(agent_id="my-llamaindex-agent")
tsv = bridge.query("screw propeller thrust calculation Bourne")

Full LangChain retriever (x402 + churn telemetry):

from unison_langchain import UnisonX402Retriever

retriever = UnisonX402Retriever(collection="unison_engineering_core", agent_id="my-rag-v1")
documents = retriever.invoke("Tesla 1891 AIEE resonant coil parameters")

---

Stream-optimized, x402-gated grounding retrievers for LangChain and CrewAI.

Drop-in LangChain retriever and CrewAI tool backed by the Unison MCP Gateway — 90,000+ vectors across 31 curated Qdrant collections covering engineering, medicine, law, finance, chemistry, and 20+ specialist domains. Data served as compact TSV streams over the x402 micro-payment protocol ($0.005 USDC/query on Base L2).

---

Why use this?

The hallucination problem is structural, not stochastic

Even at temperature=0.0, frontier models fail systematically on deep historical data. Our automated daily benchmark (see benchmarks/index.md) quantifies this precisely:

Probe	GPT-4o Claim	Primary Source	Status
Tesla operating frequency	150 kHz	Not in 1891/1892 AIEE lecture corpus	⚠️ Unverified
Typhoid cold bath threshold	103°F	102°F (Osler 1892)	❌ Protocol deviation
Year attribution	1899 notebooks	1891, 1892 published lectures	❌ Temporal conflation

Fidelity Index: 0/100 on both probes.

The token overhead is real

JSON REST APIs add structural serialization overhead that compounds at scale:

TSV  payload  [1,539 tokens]  ██████████████████░░  ← Unison format
JSON equiv.  [1,692 tokens]  ████████████████████  ← Standard REST API

8.5–9.0% token savings per payload (measured via tiktoken cl100k_base). At 1M agent queries/day → ~87,000 tokens/day eliminated.

---

Installation

# LangChain retriever only
pip install unison-langchain

# With CrewAI tool support
pip install 'unison-langchain[crewai]'

# With autonomous x402 payment settlement
pip install 'unison-langchain[payment]'

# Full install
pip install 'unison-langchain[all]'

---

Quick-start

LangChain UnisonX402Retriever

from unison_langchain import UnisonX402Retriever

retriever = UnisonX402Retriever(
    collection="unison_medical_core",
    agent_id="my-rag-chain-v1",
    k=8,
)

docs = retriever.invoke(
    "Osler 1892 typhoid fever cold bath temperature threshold Fahrenheit"
)

for doc in docs:
    print(doc.metadata["source_url"])
    print(doc.page_content[:300])
    print()

Inside a RAG chain:

from langchain.chains import RetrievalQA
from langchain_openai import ChatOpenAI
from unison_langchain import UnisonX402Retriever

qa = RetrievalQA.from_chain_type(
    llm=ChatOpenAI(model="gpt-4o", temperature=0),
    retriever=UnisonX402Retriever(collection="unison_engineering_core"),
)
result = qa.invoke({"query": "Tesla 1891 AIEE lecture resonant coil parameters"})

Auto-select collection from a hint:

retriever = UnisonX402Retriever.from_manifest_hint(
    "clinical dosing typhoid", agent_id="agent-01"
)

CrewAI UnisonGroundingTool

from crewai import Agent, Task, Crew
from unison_langchain import UnisonGroundingTool

grounding_tool = UnisonGroundingTool(
    collection="unison_engineering_core",
    agent_id="my-research-crew",
)

researcher = Agent(
    role="Senior Research Analyst",
    goal="Retrieve verified historical engineering parameters",
    backstory="You verify all technical claims against primary sources before asserting them.",
    tools=[grounding_tool],
    verbose=True,
)

---

Available Collections (25 total, 24,652 vectors)

from unison_langchain import UnisonX402Retriever
for name, desc in UnisonX402Retriever.list_collections().items():
    print(f"{name}: {desc[:80]}")

Collection	Domain	Vectors
unison_engineering_core	Tesla, Bourne, Nares, Douglas, ArXiv cs.AI	1,548
unison_medical_core	Osler, Pepper, Gray's Anatomy, Manual of Surgery	4,527
unison_manufacturing_core	Rose Machine-Shop Practice	3,374
unison_public_domain	Sun Tzu, Clausewitz, Musashi, Machiavelli, Taylor	3,700
unison_chemistry_core	Mendeleev	1,774
unison_macroeconomics_core	Smith Wealth of Nations	1,765
unison_financial_core	Mackay, SEC EDGAR 10-K FY2025/2026	1,551
unison_legal_core	Blackstone, Holmes	1,364
unison_astrophysics_core	Newton's Principia	593
…	16 more specialist domains	…

Live manifest: /.well-known/mcp-configuration

---

Payment model

• First 50 queries per agent_id — free (KV-tracked at the Cloudflare edge)
• Subsequent queries — $0.005 USDC on Base L2 via x402

For autonomous settlement, set:

export UNISON_AGENT_PRIVATE_KEY="0x..."
export UNISON_BASE_RPC_URL="https://mainnet.base.org"
export UNISON_USDC_ADDRESS="0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913"

---

Daily Benchmark Audit

An automated benchmark daemon runs daily at 03:00 UTC via GitHub Actions. Results are committed to benchmarks/index.md in this repository — a live, crawlable audit trail of GPT-4o fidelity scores and token-efficiency measurements.

View the latest: benchmarks/index.md

---

License

MIT — V18 Group