The Levels of RAG 🦜
summary
Builds a toy RAG over the conference schedule using LangChain, LangSmith and a vector DB, shows chunking and recursive splitters, adds hybrid search with BM25/TF‑IDF and Milvus via reciprocal rank fusion, and handles PDFs with PDF loaders plus table extraction (TableTransformers) — sometimes getting correct answers and sometimes wrong (payload error 8,300 vs 4,200 kg) because tables parsed poorly. Multimodal with GPT‑4o works but is expensive — $71 for 142k prompt tokens — and hybrid + chunking materially fixes many failures.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:53About Jeroen
- 1:14Why Care About RAG
- 2:15Why Use RAG
- 5:17How Does RAG Work
- 6:04Basic RAG
- 8:13Answer Generation
- 8:59Demo
- 10:41Monitoring platforms
- 12:59Chunking
- 14:13AB Testing
- 15:36Level 1 Problem
- 16:39Level 2 Problem
- 18:23Hybrid Search
- 18:51Pi Data Schedule
- 20:17Advanced Data Formats
- 22:24Debugging
- 22:59Table Transformers
- 24:11Multimodal
- 25:57Conclusion
- 27:54Questions
related talks
56mRAG: from dumb implementation to serious results
02h 47mFrom naive to advanced RAG: the complete guide
42mTake Your RAG To The Max!
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance
27mOptimizing RAG Systems in Practice: Challenges and Proven Solutions
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance
49mRAG: from dumb implementation to serious results
01h 28mFrom RAG to Riches Building Enterprise Search Systems with LLMs