From naive to advanced RAG: the complete guide
summary
Shows a concrete RAG pipeline built with LangChain4j and production deployments on Astra/Cassandra using a G‑Vector index; demonstrates specific fixes: semantic chunking (sliding‑window and parent/child), hypothetical question embedding to improve recall, query compression, metadata filtering, and reranking (BM25 + reciprocal rank fusion). Uses Gemini/Vertex and smaller embedding models (768/1536 dims), handles billion‑row stores, and runs agentic function‑calling for multistep queries. Provides GitHub examples, evaluation with recall/precision/F1 (and MRR/nDCG), and shows the Apollo 11 and Berlin demos.
AI-written summary · about
chapters (19)
- 0:00Introduction to RAG
- 3:17Understanding LLMs
- 8:46Prompt engineering basics
- 13:07LLM generation parameters
- 16:40RAG limitations and solutions
- 23:22LangChain4j overview
- 32:31Naive RAG architecture
- 37:14Ingestion and parsing
- 47:59Vectors and similarities
- 58:45Advanced chunking techniques
- 1:10:16Contextual retrieval
- 1:28:54Semantic chunking
- 1:33:36Embedding models and metrics
- 1:45:56Vector databases
- 2:00:42Query transformation
- 2:10:02Advanced retrieval techniques
- 2:23:50Evaluation and agents
- 2:32:56Agentic RAG implementation
- 2:38:40Conclusion and resources
related talks
53mBuild RAG from Scratch
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance
45mMaking semantic search & RAG real: How to build a prod-ready app
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance
27mOptimizing RAG Systems in Practice: Challenges and Proven Solutions
59mThe Evolution of RAG
21mHow RAG is Redefining AI Capabilities
50mRetrieval Augmented Generation Beyond Toy Demos