lsloan/langchain_youtube

langchain_youtube — README

A Python module to collect captions from media in YouTube and produce LangChain Document objects.

Unlike the YouTube indexer included with LangChain Community, which takes video captions and returns them as a single Document object, this module rearranges the captions into multiple Document objects, each one containing a specific number of seconds. The metadata of each Document object includes the timestamp in the video where the captions begin and the YouTube URL with the timestamp as a start time.

Installation

Download the wheel file (.whl) from the latest release on the Releases page.

pip install LangChainKaltura-0.0.1-py3-none-any.whl

Usage

Instantiate YouTubeCaptionLoader with a YouTube media URL, then invoke its load() method…

from LangChainYouTube import YouTubeCaptionLoader

# The Victors - University of Michigan 2021 Commencement
# youtu.be-style URL (watch- and embed-style URLs also supported)
mediaUrl = 'https://youtu.be/TKCMw0utiak'

captionLoader = YouTubeCaptionLoader(mediaUrl)
captionDocuments = captionLoader.load()

print('\n\n'.join(map(repr, captionDocuments)))

See the repo for example.py, a more detailed example.

Features

• Does not require a YouTube API token.
• Captions from media are reorganized into chunks. The chunk duration is configurable (chunkSeconds keyword argument), with a default of two minutes.
• The default caption languages are various English dialects, but a list of other languages may be specified (languages keyword argument).
• It returns a list of LangChain Document object(s), each containing a caption chunk and metadata.
• Caption chunks' metadata contains source URLs to the media, which includes timestamps to the specific chunk of the video.