RAG: from dumb implementation to serious results
summary
Practical RAG with LangChain4j and Gemini shows concrete fixes: implement sliding-window and recursive chunking, generate hypothetical questions or contextual summaries at ingest, use the same embedding model + in‑memory or AADB vector store, add metadata filtering and a reranker to reorder results. Demo exposes gotchas — a Berlin query returns conflicting population snippets (3.85M vs 3.75M), small local embeddings score 0.16 vs 0.72 for better models, and 30 raw hits miss combined answers — so agentic RAG plus reranking and lifecycle updates are required even with 2 million tokens context models.
AI-written summary · about
chapters (8)
- 0:00Introduction to RAG
- 3:45RAG architecture overview
- 11:39Implementing basic RAG
- 18:33Advanced ingestion techniques
- 21:03Retrieval optimization
- 39:52Refining search results
- 48:08Agentic RAG implementation
- 52:50Evaluation and conclusion
related talks
49mRAG: from dumb implementation to serious results
53mBuild RAG from Scratch
27mOptimizing RAG Systems in Practice: Challenges and Proven Solutions
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance
30mRAG: Tricks from the trenches.
42mTo RAG or not to RAG
25mAdvanced RAG Systems: Scaling Retrieval-Augmented Generation in Production
19mImproving RAG Systems - Key Challenges and Solutions