Naive RAG (embed -> retrieve -> generate) fails in production. Advanced RAG patterns: 1) Query rewriting: expand single query into 5 semantically diverse queries to improve recall - "What is RLHF?" becomes ["explain RLHF in detail", "reinforcement learning from human feedback tutorial", "how LLMs are aligned using human preferences"...]. 2) Hypothetical Document Embeddings (HyDE): generate hypothetical answer, embed it, use to retrieve real documents - often outperforms query embedding for technical topics. 3) Contextual compression: after retrieval, use LLM to extract only the relevant portions from each document rather than passing full chunks. 4) Reranking: pass top-50 retrieved chunks through CrossEncoder for relevance scoring, return top-5. 5) Corrective RAG: evaluate retrieval quality, if below threshold trigger web search to supplement knowledge. 6) Self-RAG: model decides when to retrieve (via special tokens), evaluates its own outputs for support and utility. Evaluation with RAGAS: Context Precision, Context Recall, Answer Faithfulness, Answer Relevancy - need ground truth or LLM-as-judge. Production: cache embedding computation for repeated documents, implement streaming responses, monitor retrieval quality metrics per query type.