Agentic RAG
RAG where the LLM itself drives retrieval — deciding when, what and how often to search, instead of running a simple "embed → top-k → answer" pipeline.
The model gets tools (search, fetch_url, query_db) via function calling and loops until it's satisfied. Classic RAG searches once on the user prompt; agentic RAG can decompose the question, rewrite failed queries, validate facts, follow sources and combine results across many steps.
Lifts recall and precision on complex questions (deep research, legal, scientific) but dramatically increases latency and cost — can make dozens of model calls per query.