YouTube Video, Article, and PDF Summarizer

Project Overview

This project is a web app that summarizes YouTube videos, articles, and PDFs based on a user-provided URL. It's currently in the early stages of development.

Background

I first started building this in January 2023 for a class project, about 2 months after ChatGPT's release. The goal was to create a tool for extracting information from videos to identify valuable information sources.

Initial Challenges

• Limited context windows (ChatGPT had 4k tokens)
• Inability to summarize longer texts
• Lack of web development experience

Initial Approach

• Developed a simple chunking algorithm
• Planned for webapp or Chrome extension implementation
• Used synchronous calls (inefficient)

Project Revival

Despite similar tools now existing, I'm revisiting the project to improve it and learn new skills:

• Asynchronous operations
• Performance profiling
• Retrieval-Augmented Generation
• Vector Search
• MongoDB
• Fullstack development
• Proper project structure
• Documentation
• Production-level coding practices

Roadmap

1. User provides YouTube URL (done)
2. Extract video ID (done)
3. Pull transcript if available (done)
   • If not available, transcribe with Whisper using Groq (TODO)
4. Make call to Groq LLaMA3 to provide video summary (done)
5. Chunk transcript (done)
6. Embed transcript chunks (TODO)
7. Store embedded transcript chunks in MongoDB vectorstore (pymongo) (TODO)
8. Provide user input option on screen (TODO)
   • Accept user query (TODO)
   • Embed user query (TODO)
   • Compute cosine similarity with embeddings (TODO)
   • Supply relevant embedded chunks as context to Groq LLaMA3 model (TODO)
   • Feed query to model (TODO)
9. Return model response (TODO)

Resources for Future Use

• Presidental Speeches RAG w/ Groq: https://shorturl.at/T9NkJ
• RAG w/ MongoDB Atlas: https://shorturl.at/DhY2M

Performance Improvements

Performance Optimization

All speedups as measured from base.

Version	Runtime (s)	Function Calls	Speedup
Base (Sync GPT-3.5)	25.235	242367	-
Async GPT-3.5	9.198	73185	2.74x
Async GPT-4o Mini	8.999	54709	2.8x
Async Llama-70B (Groq)	4.844	53760	5.63x

Base Functionality (Synchronous with GPT-3.5-turbo)

• Runtime: 25+ seconds
• 242k+ function calls
• 95% of runtime in SSLSocket.read()

Asynchronous Version

• Runtime: < 9.2 seconds
• 74k function calls
• 2.74x speedup (63.55% reduction)

Llama 70B via Groq (Asynchronous)

• Runtime: 4.84 seconds
• 53.7k function calls
• 5.63x total speedup (82.23% reduction from base)

Recent Updates

August 10

• Restructured directories to add frontend, backend, and testing dirs
• Added testing for video.py
• Caught a bug in video.get_transcript() - default language was 'en' and so no transcript was getting pulled. Rewrote so tries specificed language, then tries English, and then falls back on first available transcript

August 9

• Improved error handling and imports in video.py and summary files
• Optimized code: better docstrings, naming, and separation of concerns
• Created sync_summary.py and async_summary.py
• Integrated Llama-70B-8192 via Groq API
• Switched from GPT-3.5-turbo to GPT-4-mini
• Implemented profiler for benchmarking

August 8

• Restructured directory
• Created basic HTML and CSS for future use

Future Improvements

• Increase chunk size for larger context windows
• Implement adaptive chunking based on model context size
• Profile smaller Groq model & test quality
• Evaluate output quality of different models

Project Status

The project is functional but still has significant room for improvement and expansion.