Build RAG from Scratch
summary
Live-coded a working RAG pipeline that starts with a 2,938-dim bag-of-words vector built from ~100 scraped talk titles and descriptions, compares queries using cosine similarity, then replaces the crude vectors with real embeddings (OpenAI text-embedding-3-small 1536-dim) and a vector index in Astra DB vector index. The prototype plugs those results into Gemini 1.5 Flash (and can swap in a 7B LLM), boosts the correct-match score from ~0.24 to ~0.83, and demonstrably answers event-specific queries.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 3:20LLMs need augmentation
- 5:06What is RAG
- 7:56Natural language search
- 9:02Vector embeddings example
- 16:24Cosine similarity
- 27:47Bag-of-words limits
- 30:18Embeddings and vector DBs
- 33:53Why vector databases are essential
- 37:48RAG demo with vector DB
- 40:38Chunking text
- 43:27Storing data and models
- 45:16Advanced models like DeepSeek
- 48:25Retrieval is key
- 49:28ColBERT and Graph RAG
- 51:16Vector search beyond LLMs
- 51:58RAG future and context
related talks
44mBuild RAG from Scratch
56mRAG: from dumb implementation to serious results
02h 47mFrom naive to advanced RAG: the complete guide
49mRAG: from dumb implementation to serious results
01h 35mTutorial: Build AI-Supercharged RAG Apps with a Vector Database
21mHow RAG is Redefining AI Capabilities
25mSet Up an Open Source RAG in 30 Minutes
30mRAG: Tricks from the trenches.