Что такое RAG (Retrieval-Augmented Generation)?

RAG solves a real LLM limitation: training data has a cutoff date, and most companies' internal documents are never in training data at all. Instead of retraining the model (slow, expensive), RAG keeps your documents in a vector database (Pinecone, Weaviate, Postgres pgvector), embeds the user's question into the same vector space, finds the most relevant chunks, and feeds those chunks to the LLM as additional context. The model then answers based on those chunks. Result: an LLM that can answer questions about your specific docs, with citations, without expensive retraining. RAG dominates enterprise AI in 2026 — almost every 'chat with your PDFs' or internal-knowledge bot is RAG under the hood. The hard parts: choosing good chunk sizes, getting embeddings to actually match user intent, and managing the latency of retrieve-then-generate flows.