langchain-finic

Community-maintained LangChain integration for Finic Grayson, by Suprep Labs. Typed decisions, LCEL composition, LangGraph nodes, structured tools, and optional agent middleware. This alpha is not endorsed by Finic or LangChain.

Documentation · Source · PyPI · MIT license

Install

Python 3.11 or newer. Install the alpha release:

pip install langchain-finic==0.1.0a4
export FINIC_API_KEY=your_key
# Optional; defaults to grayson-1:
export FINIC_MODEL=grayson-1

Make a typed decision

from langchain_finic import Choice, DecisionRequest, FinicClassifier, Noul

request = DecisionRequest(
    state={"events": ["New device", "Transfer to an unfamiliar recipient"]},
    questions={
        "takeover": Noul(instructions="Does this evidence suggest account takeover?"),
        "route": Choice(
            instructions="Choose a review queue.",
            criteria={"standard": "Routine review", "urgent": "Urgent investigation"},
        ),
    },
)

with FinicClassifier() as classifier:
    decision = classifier.invoke(request)
    print(decision.nouls["takeover"].noul)
    print(decision.choices["route"].choice)
    print(decision.request_id)

Noul, Choice, MultiChoice, and Score have typed answers. Responses retain probabilities, usage, and request metadata. Finic evaluates decisions; it does not generate chat messages or text tokens.

Compose with LangChain and LangGraph

# request is defined above.
with FinicClassifier() as classifier:
    route_chain = classifier | (lambda result: result.choices["route"].choice)
    route = route_chain.invoke(request, config={"tags": ["risk-review"]})
    tool = classifier.as_tool()
    print(tool.name)  # finic_decide; pass it to create_agent or ToolNode

import asyncio

async def review():
    async with FinicClassifier() as classifier:
        return await classifier.abatch([request], config={"max_concurrency": 1})

results = asyncio.run(review())

Batching makes individual API requests; concurrency is controlled by Runnable config. Streaming yields one complete decision. A chat-provider model can call the Finic tool through create_agent; the Finic integration itself is a decision Runnable.

Optional middleware

pip install 'langchain-finic[middleware]==0.1.0a4'

ModelRouterMiddleware selects among host chat models. AutoModeMiddleware screens specified tools before execution. See the complete middleware example and configuration guide. Tool screening is probabilistic; application authorization remains necessary.

Configuration and data

Use FINIC_API_KEY, FINIC_MODEL, FINIC_TIMEOUT_SECONDS, and FINIC_MAX_ATTEMPTS, or constructor options. The library does not load .env. The host controls optional LangSmith tracing, which can record request data and responses. Default timeout is 60 seconds, with up to three attempts; retries can incur additional charges. Use context managers to reuse and close owned HTTP clients. Injected clients remain caller-owned. Read the usage guide for precedence, serialization, resource lifetimes, and error handling.

Development

uv sync --locked --group dev --extra middleware
make check

Default tests use mocked HTTP with network disabled. Opt-in live tests spend Finic credits. See contributing and security reporting.

References: Finic API, LangChain integration guidelines.