Advanced RAG Systems: Scaling Retrieval-Augmented Generation in Production
summary
Covers concrete recipes: use BM25+vector hybrid retrieval to return ~20–50 chunks, then run a cross‑encoder reranker on ~20 candidates to reduce to top 5–10 chunks for the final prompt; apply HyDE hypothetical answers and reciprocal rank fusion for multi‑query boosts. Describes Graph RAG and Light RAG tradeoffs (single canonical nodes, incremental updates), multimodal embedding for videos/images, and production metrics via DeepEval and DAG evaluation pipelines, with notes on latency and token savings.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:11RAG in technical interviews
- 3:38Chunking strategies
- 6:23Embeddings and multimodal data
- 8:33Hybrid search BM25 and vectors
- 11:03Graph RAG and knowledge graphs
- 12:25LightRAG and incremental graphs
- 13:54Open source for job seekers
- 14:36Query rewriting techniques
- 16:42Hypothetical answers and embeddings
- 17:24Reranking and prompt engineering
- 19:33Evaluating RAG with DeepEval
- 20:38Adaptive RAG and agents
- 21:39Live demo and Q&A
related talks
50mRetrieval Augmented Generation Beyond Toy Demos
21mHow RAG is Redefining AI Capabilities
51mDeveloping a Custom RAG System
56mRAG: from dumb implementation to serious results
15mRAG Beyond Chatbots
27mOptimizing RAG Systems in Practice: Challenges and Proven Solutions
28mImproving responses in generative AI applications
39mRAGs to Riches: How Our Content Affects Retrieval Augmented Generation