An Introduction to Retrieval Augmented Generation
summary
Creates a working Retrieval Augmented Generation pipeline that pulls YouTube videos transcripts and sampled frames, chunks text (token and recursive strategies), builds embeddings with text embedding three small and CLIP for images, stores vectors in LanceDB, and persists a multimodal index via llama index. Notebook shows embedding sizes (1,536 token vectors reduced for storage), chunk counts changing (446 → 496), k=6 retrievals, and demo queries; code runs end-to-end with occasional OpenAI API key/setup issues but yields grounded, multimodal retrieval results.
AI-written summary · about
chapters (9)
- 0:00Introduction to Retrieval Augmented Generation
- 9:00Why use RAG
- 17:40Extracting YouTube info
- 29:41Chunking text/video for RAG
- 38:47Generating embeddings
- 46:25Storing embeddings
- 57:41Building RAG pipeline
- 1:00:19Testing RAG system
- 1:13:04Live demo & Q&A
related talks
50mRetrieval Augmented Generation Beyond Toy Demos
31mRetrieval Augmented Generation
48mAn Introduction to Retrieval Augmented Generation
01h 00mLLMs gone wild
27mUnlocking the Potential of Large Language Models
25mAdvanced RAG Systems: Scaling Retrieval-Augmented Generation in Production
48mGoing Beyond Prompt Engineering
33mVector Databases for GenAI applications