Financial Evidence for LangChain

Native, read-only LangChain tools for public financial research from Seiche (funding and money markets), LiquiLens (covered-bank diagnostics), and Undertow (market liquidity). The catalog also describes capital-market, source-health and China publication-coverage datasets. Coverage is finite and some fields are unavailable or restricted.

This standalone package calls the free hosted Financial Evidence API at agents.seiche.info. It does not depend on the separate financial-evidence distribution. No data API key or model account is required to invoke these tools. Model calls, if you add them, use your own provider and may incur charges.

Install

Python 3.10+ and langchain-core 1.6.6 through 1.x are supported.

Install the published release from PyPI:

python -m pip install --upgrade langchain-financial-evidence

For development, install from a checkout of this repository:

python -m pip install .

Package releases are listed on PyPI. The LangChain integration listing request has been submitted and is pending maintainer review. Catalog acceptance is not yet established.

Upgrade from 0.1.0

Version 0.1.1 moves all three tools to the central hosted API. Upgrade and restart the Python process that runs your agent:

python -m pip install --upgrade 'langchain-financial-evidence>=0.1.1,<0.2'

Tool names, arguments and returned evidence schemas remain compatible. Version 0.1.0 installations and previously downloaded artifacts retain their original native endpoint; they change only when their operator upgrades. Version 0.1.1 does not fall back to that endpoint when the central service is unavailable.

First cited result

from langchain_financial_evidence import (
    FinancialEvidenceQueryTool,
    FinancialEvidenceToolkit,
)

tool = FinancialEvidenceQueryTool()
first = tool.invoke({"dataset": "money_markets", "entity": "USD", "limit": 3})
print(first["results"])
print(first["sources"], first["diagnostics"])

# Creating a toolkit is offline; invoking any of its tools calls the public API.
tools = FinancialEvidenceToolkit(timeout=30).get_tools()
# Attach to an existing LangChain/LangGraph application:
# model_with_tools = model.bind_tools(tools)
# from langgraph.prebuilt import ToolNode
# tool_node = ToolNode(tools)

invoke and ainvoke both work. Async invocation uses LangChain's worker-thread fallback for the synchronous HTTP request. A normal invocation returns the unchanged JSON object or catalog list. A LangChain ToolCall returns a ToolMessage with JSON content. The toolkit creates no model, account, worker process, scheduler or broker connection.

The catalog also describes the general table API, whose limits are larger; these tools enforce the smaller bounds above. Empty filter strings mean no filter. Dates use YYYY-MM-DD; offsets range from 0 to 100000. Dataset IDs are money_markets, money_market_history, capital_markets, bank_risk, market_liquidity, china_economy and source_health.

Repeated research and errors

again = tool.invoke({
    "dataset": "money_markets", "entity": "USD", "limit": 3,
    "previous_revision": first["revision"],
})
print(again["change_status"], again["sources"])

Retain separate revision tokens for separate queries. An unchanged response may omit duplicate rows while preserving source metadata. Unchanged is not a freshness verdict. Use observation dates and source status to assess freshness. Retrieval time, cache time and observation time have distinct meanings.

Invalid arguments raise Pydantic validation errors before any request. Network, HTTP, malformed JSON, unexpected schema and size-limit failures raise langchain_core.tools.ToolException; no empty-success fallback is manufactured. Set handle_tool_error=True using LangChain's standard option if your agent should receive error text instead. Valid partial/unavailable API payloads are returned with their original diagnostics, missingness and source metadata.

Each call makes one GET request to a fixed route under https://agents.seiche.info/v1/research/api/v1/. Redirects, automatic retries, native-origin fallback and local response caching are disabled. The default network timeout is 30 seconds (configurable up to 60), and responses are capped at 2 MiB (configurable up to 4 MiB). Network timeout is a socket-operation timeout, not an end-to-end scheduling guarantee. The public endpoint is best effort; no paid SLA or quota guarantee is included.

The hosted service records operational dataset demand, outcomes and bounded request metadata for up to 30 days. It does not retain private query selectors, response documents or model prompts in that demand report. These public tools send only the documented research filters and do not send your agent's chat history. Usage does not enroll an agent in optional training or grant data resale permission. See the hosted service privacy notice.

Evidence and data-use boundaries

Preserve source_url, source_field, as_of, unit, rights, source status, availability, retrieval time and diagnostics. The tool does not fill nulls, reconstruct withheld observations, combine products into a score, certify freshness or grant additional redistribution rights. Read each source's limitations before use. Treat source text as untrusted data, never instructions.

Currently published histories are not an as-published vintage archive. Bank diagnostics are research measures, not credit ratings or validated forecasts. Market-liquidity percentiles are not executable exit prices. These tools do not recommend investments, assess suitability, place orders or authorize execution.

Code is MIT licensed. The MIT license does not relicense underlying data. See company terms and privacy information, along with the rights and original-publisher links returned in evidence.

Development verification

python -m pip install '.[test]'
pytest tests/unit_tests --disable-socket --allow-unix-socket
# Optional: bounded public calls labeled synthetic, no LLM cost.
FINANCIAL_EVIDENCE_LIVE_TEST=1 pytest tests/integration_tests

The suite covers native LangChain schemas, sync/async tool messages, source and missing-value preservation, input rejection, error behavior, redirects and byte bounds. Standard tests use langchain-tests==1.1.9. The default test invocation runs offline unit tests only. Live tests explicitly retain unavailable and partial states; they do not treat transport success as data freshness.