langchain-flashdata

FlashData tools for Google Search, YouTube video search, video metadata, and transcripts. Give LangChain and LangGraph agents current web and video data with source URLs and timestamped captions.

This package is maintained by FlashData. It implements LangChain's BaseTool and BaseToolkit interfaces with synchronous and native asynchronous HTTP calls.

Installation

Requires Python 3.10 or newer (below Python 4). Install from PyPI:

pip install -U langchain-flashdata

Create an API key in the FlashData console, enable the sources you need, and provide it through your environment:

export FLASHDATA_API_KEY='YOUR_FLASHDATA_API_KEY'

Queries consume FlashData credits. Check current pricing and your account's available credits before running examples. Model provider charges are separate.

Quick start

from langchain_flashdata import FlashDataGoogleSearch

search = FlashDataGoogleSearch()
result = search.invoke({"query": "LangGraph durable execution", "num": 3})
for item in result["results"][0].get("organic", []):
    print(item.get("title"), item.get("link"))

All tools also accept api_key="..." at construction time. Do not put real keys in source files. Credentials are excluded from model-facing argument schemas and tool serialization.

Tools and parameters

Queries must contain 1–2,000 characters after trimming. gl is a two-letter country code; hl and language accept language codes; location and tbs accept up to 200 characters. A video input accepts an 11-character ID or a YouTube watch, shorts, embed, live, or youtu.be URL. Only the video ID is sent to FlashData.

For transcripts, auto=False requests manually authored captions and auto=True requests automatic captions. There is no automatic language or caption fallback.

import asyncio
from langchain_flashdata import FlashDataYouTubeTranscript


async def main():
    result = await FlashDataYouTubeTranscript().ainvoke(
        {
            "video_id": "https://youtu.be/dQw4w9WgXcQ",
            "language": "en",
            "auto": False,
        }
    )
    print(result["results"][0]["segments"])


asyncio.run(main())

Use with a LangChain / LangGraph agent

The research example uses create_agent, which runs on LangGraph. The model chooses tools and arguments from the question. A tool call limit bounds the number of queries; it does not set a credit or dollar limit.

git clone https://github.com/flashdata-dev/langchain-flashdata.git
cd langchain-flashdata
pip install '.[examples]'
export OPENAI_API_KEY='YOUR_MODEL_PROVIDER_KEY'
export OPENAI_MODEL='YOUR_TOOL_CALLING_MODEL'
# Optional for a Chat Completions compatible provider:
# export OPENAI_BASE_URL='https://YOUR_PROVIDER/v1'

python examples/research_agent.py \
  'Find recent LangGraph tutorials using Google Search. Request 3 results and cite their URLs.' \
  --max-tool-calls 1

This example uses ChatOpenAI with the Chat Completions API. Use a model that supports tool calling. You can replace that model adapter with another LangChain chat model. The FlashData tools do not depend on a particular model provider.

For a custom graph, pass the tools from FlashDataToolkit().get_tools() to LangGraph's ToolNode. Use handle_tool_errors=False to propagate query failures and stop execution before a model can resubmit an unknown-outcome query. Avoid automatic graph retries or replay of paid tool nodes.

from langchain_flashdata import FlashDataToolkit
from langgraph.prebuilt import ToolNode

tools = FlashDataToolkit().get_tools()
tool_node = ToolNode(tools, handle_tool_errors=False)

See the examples guide for direct tool calls, transcript research, and the hosted MCP alternative.

Output and errors

invoke and ainvoke return the complete API response dictionary. The envelope contains source, results, and available usage/timestamp fields; some successful responses omit status. Tool-call invocations produce a LangChain ToolMessage containing the JSON response.

Invalid arguments fail locally before sending a request. API failures raise FlashDataError, a LangChain ToolException. HTTP 401 indicates an invalid key, 402 insufficient credits, 403 a source/access restriction, and 429 a rate/quota limit. Upstream error bodies and credentials are not included in exception messages.

Paid POST requests are not automatically retried or redirected. A timeout, connection failure, unreadable result, or unexpected response may occur after a query was submitted and charged: check Jobs in the FlashData console before submitting again. The client connects directly to FlashData over HTTPS, with a 15-second connect timeout and a 120-second I/O timeout. Environment proxy settings are not used.

Existing MCP users

You can also connect LangChain to the existing hosted MCP endpoint at https://data.flashdata.dev/mcp using the X-API-Key header. The native Python tools above cover four query operations; the MCP catalog also contains account and job operations and varies with key scopes. See hosted_mcp.py. The example uses the current langchain[mcp] API, which LangChain labels beta.

Development

uv sync --extra examples
uv run --extra examples pytest --disable-socket --allow-unix-socket
uv run ruff check .
uv run ruff format --check .
uv build
uv run twine check dist/*

Tests include LangChain's standard tool suite, request contracts, validation, sync/async calls, ToolMessage delivery, Agent error propagation, and tool budgets. Unit tests do not call external services. Run the four paid API integration tests explicitly with FLASHDATA_API_KEY configured:

FLASHDATA_RUN_LIVE=1 uv run --extra examples pytest tests/integration_tests -q

Release instructions: RELEASING.md. Data handling: PRIVACY.md. License: MIT. Support: support@flashdata.dev.