iflow-search

Python SDK for iFlow Search (心流搜索) — a search API that exposes web search, image search, and web-page fetching with AI-friendly structured output.

• Product: https://platform.iflow.cn/
• API docs: https://platform.iflow.cn/docs/
• Status: stable (0.1.0), published on PyPI.

This is the framework-agnostic core SDK: it has zero LangChain / MCP / FastAPI dependencies. The MCP adapter (iflow-search-mcp), the LangChain adapter (iflow-search-langchain), and the FastAPI / OpenAPI adapter (iflow-search-openapi) all ship from the same repository and depend on this package.

Install

pip install iflow-search

For local development:

git clone https://github.com/zhengyanglsun/iflow-search-py.git
cd iflow-search-py/packages/iflow-search
python -m pip install -e ".[dev]"

Quick start (sync)

import os
from iflow_search import IFlowSearchClient

client = IFlowSearchClient(api_key=os.environ["IFLOW_API_KEY"])

result = client.web_search(query="flash attention", count=5)
for r in result.results:
    print(r.title, r.url)

images = client.image_search(query="iflow logo")
for i in images.images:
    print(i.image_url, i.source_url)

page = client.web_fetch(url="https://platform.iflow.cn/docs/")
print(page.title, len(page.content), "chars")

Quick start (async)

import asyncio, os
from iflow_search import AsyncIFlowSearchClient

async def main() -> None:
    async with AsyncIFlowSearchClient(api_key=os.environ["IFLOW_API_KEY"]) as client:
        result = await client.web_search(query="flash attention", count=5)
        for r in result.results:
            print(r.title, r.url)

asyncio.run(main())

Configuration

Parameter	Default	Notes
api_key	os.environ["IFLOW_API_KEY"]	Required. Falls back to env var; raises IFlowConfigError if neither is set.
source	"python"	Sent as IFlow-Source header. Adapter packages override this.
integration_name	"iflow-search"	Sent as IFlow-Integration header.
integration_version	installed package version	Sent as IFlow-Integration-Version header.
base_url	https://platform.iflow.cn	Useful for staging/proxy environments.
timeout	30.0 seconds	Applied to each individual request.
http_client	new httpx.Client (or AsyncClient)	Inject one for connection pooling or testing (e.g. via httpx.MockTransport).

Attribution headers

Every outbound request carries:

Authorization:             Bearer <api-key>
Content-Type:              application/json
Accept:                    application/json
IFlow-Source:              python   (configurable)
IFlow-Integration:         iflow-search   (configurable)
IFlow-Integration-Version: <installed package version>
User-Agent:                iflow-search/<version>

These are built only by the SDK; constructor arguments let adapter packages set their own source / integration_name / integration_version. There is no API for users to override Authorization, Content-Type, Accept, or User-Agent directly.

Errors

The SDK raises a typed exception hierarchy — never tuples, dicts, or sentinel values:

IFlowError
├── IFlowConfigError                 # missing api_key, invalid attribution
├── IFlowValidationError             # bad client-side input
├── IFlowAuthError                   # HTTP 401/403, business code 90402
├── IFlowRateLimitError              # HTTP 429, business code 40303
├── IFlowInsufficientCreditsError    # business code 60400
├── IFlowAPIError                    # HTTP 5xx, non-JSON 2xx, other non-2xx
├── IFlowBusinessError               # success=false with any other code
├── IFlowTimeoutError                # SDK-initiated timeout
└── IFlowNetworkError                # DNS / connection / TLS errors

Every exception carries code (stable string), message, request ({method, url, endpoint}), and response_body_truncated (first 500 chars when applicable). Switch on code rather than class identity if you want a stable contract across SDK versions.

from iflow_search import IFlowSearchClient, IFlowError, IFlowRateLimitError

# Reads IFLOW_API_KEY from the environment.
client = IFlowSearchClient()

try:
    result = client.web_search(query="latest LLM benchmarks", count=3)
except IFlowRateLimitError:
    print("Rate limit exceeded")
except IFlowError as exc:
    print(f"iFlow error: {exc.code} {exc.message}")

asyncio.CancelledError is not wrapped — it propagates as itself so cooperative cancellation keeps working.

Field renames

The wire format uses a few awkward field names. The SDK renames them:

Request	Wire	Python
web/image	keywords	query
web/image	num	count

Response	Wire	Python
web	link	url (on WebSearchResult)
image	url	image_url
image	refUrl	source_url
web_fetch	fromCache	from_cache
all	—	took_ms (measured client-side)

The raw envelope is preserved on response.raw for callers that need fields the SDK did not model.

Security

• Never commit API keys. Use environment variables (IFLOW_API_KEY) or your platform's secret manager.
• The SDK does not auto-load .env files and does not read the key from any filesystem path other than the process environment.
• Tests in this repository use fake keys (literal test-key) and httpx.MockTransport — no real API is contacted, ever.

Local development

From packages/iflow-search/:

python -m pytest -q                    # offline test suite
python -m ruff check .                 # lint
python -m mypy src/iflow_search        # strict typecheck
python -m build                        # build sdist + wheel into dist/

Real-API smoke

A separate opt-in script exercises all three endpoints against the live API:

export IFLOW_API_KEY="your-api-key"
export IFLOW_SMOKE=1
python scripts/smoke_real_api.py

The script:

• Is opt-in — without IFLOW_SMOKE=1 it refuses to call the live API.
• Reads IFLOW_API_KEY from the environment only — never from disk.
• Redacts the key in all log output.
• Does not write any file.

Adapter packages

This package is the core. Adapter packages depend on it and live in the same repository:

• iflow-search-mcp — published on PyPI. MCP stdio server for Claude Code, Claude Desktop, Hermes, OpenCode, and other MCP-capable hosts. Install with pip install --pre iflow-search-mcp.
• iflow-search-langchain — published on PyPI. LangChain (and LangGraph) tools; usable from CrewAI via a one-line BaseTool subclass. Install with pip install --pre iflow-search-langchain.
• iflow-search-openapi — published on PyPI. FastAPI / OpenAPI 3.1 tool server for Open WebUI, Coze, and other OpenAPI tool catalogues. Install with pip install --pre iflow-search-openapi.

License

MIT