Vidooo

A RAG application that lets you chat with YouTube videos using their transcripts. Ask questions and get answers with timestamp references.

Features

• Transcript Processing: Fetches YouTube transcripts with timestamps
• Q&A System: Ask questions about video content
• Timestamp References: Get specific timestamps for answers
• Vector Storage: ChromaDB for efficient retrieval
• LLM Integration: Powered by Google Gemini 2.5 Flash

Tech Stack

• Python 3.12+ | LangChain | ChromaDB | Google Generative AI | YouTube Transcript API

Quick Start

1. Install

   git clone <repo>
   cd youtube-summarizer
   uv pip install -e .

2. Configure

   GOOGLE_API_KEY=your_key_here
   COLLECTION_NAME=video_chunks
   EMBEDDING_MODEL=text-embedding-004

3. Use

   # Ingest video (update URL in scripts/ingest_vid.py)
   uv run python scripts/ingest_vid.py
   
   # Start chatting
   uv run python rag_pipeline/qa_chains.py

Project Structure

youtube-summarizer/
├── rag_pipeline/     # Core RAG components
├── scripts/         # Utility scripts
├── chroma_db/       # Vector storage
└── pyproject.toml   # Configuration

TODO / Development Tracker

In Progress

• Chunking Strategy: Merge small transcript chunks for better context
• Hallucination Reduction: Improve prompt engineering and retrieval

Planned Features

• Web Extension: Browser extension for YouTube integration
• Backend API: FastAPI server for web interface
• Hybrid Embeddigs: Try VLMs along with transcripts
• Caching System: Cache embeddings to avoid re-computation
• Error Handling: Better error messages and recovery
• Testing Suite: Unit tests for all components
• Documentation: API docs and usage examples

Completed

• Core RAG pipeline
• YouTube transcript extraction
• Vector storage with ChromaDB
• Q&A system with timestamps
• Language-aware processing
• Debug functionality
• Chat history support

Debug

Use debug mode to inspect retrieval:

# In chat interface, type: debug
# Then enter query to see retrieved documents

License

MIT License

---