powered by Javascript, Python, Flask, and OpenAI GPT-3 models
sb_demo_2.mp4
This project's goal is to package the functionality of OpenAI's language processing models into an accessible and lightweight browser application. This project follows a client server design pattern, using an external flask server for API requests and data parsing.
-
Make video content more straightforward by extracting & rendering it as text within the browser.
-
Format by topic and package content into toggleable sections, containing a heading, body, and if applicable, Markdown formatting.
-
Identify topics' specific mention during a video and generate timestamps for quick navigation.
-
Enable keyword and important phrase searching through rendered text.
-
Package and export text as PDFs for later reference and local storage.
This project is currently in development.
- Ongoing tasks...
- Flask server's handling for OpenAI errors
- Client's handling for Flask server errors
- Overflow text leaving blank space in PDF exports
- Parsing YT desc sections
- PDF export options
- Reversing sidepanel iframe implementation
- Minimize sidepanel
- Bundling multiple OpenAI requests for longer videos
- This extension has not been deployed onto the Chrome Web app store
- In addition, the web server has not been deployed (runs on localhost)
This project is best suited for videos 8-18 minutes long, and reaches inaccuracies beyond 30 minutes.
- OpenAI proposes a 4000 word token limit per request
- Videos longer are shortened in order to fit inside the 4000 word limit
- Shortening does not mean an abrupt cutoff, instead removing unnecessary phrases. However, excessive length may lead shortening to cause inaccuracies.
- Until multiple request functionality is finished.
Timestamps are liable to inaccuracy
- Current ID involves matching headings with phrases within the video's transcript
- Headings can be mismatched with unrelated sections
- Can be alleviated by parsing YT sections in descriptions
Exporting PDFs with Markdown is currently not supported.
- While the browser is able to render Markdown thanks to "md-block.js", the current PDF exporting does not recognize the syntax and outputs as plain text
This project uses GPT-3 modes in parsing Youtube video content, and is liable to misrepresented extracted info.
- As opposed to summarization, this project attempts to cover video content in bulk.
- As a result, it is poor at detecting conveyed sarcasm and/or nuance.
Pull requests are welcomed. For major changes, please open an issue first to discuss what you would like to change.