langchain/docs/ecosystem/huggingface.md

Hugging Face

This page covers how to use the Hugging Face ecosystem (including the Hugging Face Hub) within LangChain. It is broken into two parts: installation and setup, and then references to specific Hugging Face wrappers.

Installation and Setup

If you want to work with the Hugging Face Hub:

• Install the Hub client library with pip install huggingface_hub
• Create a Hugging Face account (it's free!)
• Create an access token and set it as an environment variable (HUGGINGFACEHUB_API_TOKEN)

If you want work with the Hugging Face Python libraries:

• Install pip install transformers for working with models and tokenizers
• Install pip install datasets for working with datasets

Wrappers

LLM

There exists two Hugging Face LLM wrappers, one for a local pipeline and one for a model hosted on Hugging Face Hub. Note that these wrappers only work for models that support the following tasks: text2text-generation, text-generation

To use the local pipeline wrapper:

from langchain.llms import HuggingFacePipeline

To use a the wrapper for a model hosted on Hugging Face Hub:

from langchain.llms import HuggingFaceHub

For a more detailed walkthrough of the Hugging Face Hub wrapper, see this notebook

Embeddings

There exists two Hugging Face Embeddings wrappers, one for a local model and one for a model hosted on Hugging Face Hub. Note that these wrappers only work for sentence-transformers models.

To use the local pipeline wrapper:

from langchain.embeddings import HuggingFaceEmbeddings

To use a the wrapper for a model hosted on Hugging Face Hub:

from langchain.embeddings import HuggingFaceHubEmbeddings

For a more detailed walkthrough of this, see this notebook

Tokenizer

There are several places you can use tokenizers available through the transformers package. By default, it is used to count tokens for all LLMs.

You can also use it to count tokens when splitting documents with

from langchain.text_splitter import CharacterTextSplitter
CharacterTextSplitter.from_huggingface_tokenizer(...)

For a more detailed walkthrough of this, see this notebook

Datasets

The Hugging Face Hub has lots of great datasets that can be used to evaluate your LLM chains.

For a detailed walkthrough of how to use them to do so, see this notebook