retrieval augmentation

Retrieval Augmented Generation (RAG) enhances LLM outputs with retrieved external information for accurate, current responses. **Core pattern**: User query → retrieve relevant documents from knowledge base → inject documents into LLM context → generate grounded response. **Why RAG?**: LLMs have knowledge cutoffs, may hallucinate, can't access private data. RAG provides current, verified, domain-specific information. **Components**: Document store (chunked, embedded), vector database, retriever, reranker (optional), generator (LLM). **Pipeline variations**: Naive RAG, advanced RAG (query rewriting, reranking), modular RAG (customizable stages). **Quality factors**: Retrieval precision/recall, chunk quality, context window usage, generation faithfulness. **Applications**: Enterprise search, customer support, research assistants, code assistants, domain experts. **Frameworks**: LangChain, LlamaIndex, Haystack, custom implementations. **Challenges**: Retrieval failures, context length limits, maintaining freshness, evaluation. **Optimization areas**: Embedding models, chunking strategy, retrieval algorithms, reranking, prompt engineering. RAG is the dominant pattern for production LLM applications requiring accuracy.