RAG — Retrieval-Augmented Generation

Status: 🚧 Coming soon — dedicated chapters are being written.

RAG is the technique of grounding an LLM's response in retrieved, up-to-date documents instead of relying on what's baked into the model's weights. It's how production LLM apps stay accurate, current, and citable.

What this section will cover

• Why RAG: hallucination, freshness, citations, cost
• The RAG loop: chunk → embed → store → retrieve → rerank → generate
• Chunking strategies (fixed, semantic, recursive, document-aware)
• Embedding models (OpenAI, Cohere, HuggingFace, local)
• Vector stores (FAISS, Chroma, pgvector, Pinecone, Weaviate)
• Retrievers (similarity, MMR, hybrid BM25 + dense, multi-query)
• Reranking (cross-encoders, Cohere Rerank)
• Evaluation (faithfulness, context relevance, answer relevance)
• Production patterns: caching, observability, fallback

Currently available — related material

RAG components are already covered in the LangChain section:

• LangChain → Document Loaders
• LangChain → Text Splitters
• LangChain → Vector Stores
• LangChain → Retrievers
• LangChain → RAG

A dedicated RAG track (end-to-end pipelines, evaluation, production tips) is in progress.