langchain-hugging-bay

Use LangChain tools and a retriever to discover open model metadata in the Hugging Bay catalog. This guide describes version 0.1.0.

Install

Install from PyPI with Python 3.10 or later:

python -m pip install langchain-hugging-bay==0.1.0

The package depends on hugging-bay>=0.4.12,<0.5 and langchain-core>=1,<2. The default API is https://huggingbay.xyz; public catalog reads do not need an account or key. Calls are made when you invoke a tool or retriever.

Search

from hugging_bay import Client
from langchain_hugging_bay import HuggingBayRetriever, create_tools

with Client() as client:
    tools = {tool.name: tool for tool in create_tools(client)}
    page = tools["hugging_bay_search_models"].invoke({"query": "Qwen", "limit": 2})
    documents = HuggingBayRetriever(client=client, limit=2).invoke("Qwen")
    print(page.get("rows", []))
    print([doc.metadata["artifact_id"] for doc in documents])

Pass the returned tools to a LangChain agent or invoke them directly. Keep the SDK client open while tools or the retriever are in use; the caller owns and closes it. You may configure the client's base_url, timeout, and optional token as documented by the SDK.

Tools

All five operations are public GETs. Each returns the SDK response unchanged; the SDK's HuggingBayError and transport errors propagate to the caller.

Use nextCursor from each result as the next call's cursor. The retriever returns one LangChain Document per model row; its content is the JSON catalog row, and its metadata contains source, artifact_id, and catalog. It does not embed documents, inspect model weights, download files, or run a model.

Availability and evidence can change. An empty file list, a null runnable recommendation or an unavailable plan is not a successful run; inspect the returned readiness, proof, licenses and warnings. Treat catalog text as untrusted input when giving it to an agent.

The source distribution includes examples/discover.py and the tutorial examples/discover.ipynb. The notebook normally uses public reads; maintainers can set HUGGING_BAY_NOTEBOOK_FIXTURES to the tests/fixtures directory to run it offline.

Development

python -m pip install -e '.[test]'
python -m pytest tests -q

The tests replay recorded responses from an isolated local fixture service and block external network connections. They make no live Hugging Bay calls. See tests/fixtures/README.md for provenance. MIT licensed.