How does agentic RAG differ from single shot RAG, and what does putting retrieval inside the loop enable?

**Single shot RAG** is the simple version. You embed the user's question, find the top matching chunks from a vector store, paste them into the prompt, and ask the model to answer. Retrieval happens exactly once, before generation starts. That works well when the answer sits in one place. It breaks when the first search misses, or when the question needs two lookups that depend on each other. **Agentic RAG** fixes this by turning retrieval into a tool the agent can call inside its loop. Instead of one fixed search, the model reasons, decides it needs information, issues a search, reads the result, and then decides whether it needs another search. This unlocks three things. The agent can do multi-hop lookups, where the first result tells it what to search for next. It can rewrite a vague query into a sharper one after seeing weak results. And it can check whether the retrieved text actually answers the question before committing. The price is real. Every extra retrieval is another model call, so agentic RAG is slower and more expensive than the single shot version.

**Single shot RAG** retrieves once. You embed the question, pull the top-k chunks, inject them, and generate. The retrieval policy is fixed and blind to what the model later realises it needs. Agentic RAG instead exposes retrieval as a tool call inside the agent loop. The model decides when to retrieve, what query to issue, and how many times to search, all conditioned on its current reasoning state. Retrieval becomes a decision, not a preprocessing step. That shift enables three capabilities single shot RAG cannot offer. Multi-hop retrieval means the first result reveals an entity, and the agent issues a follow up query that depends on it. Query rewriting means weak or too broad chunks trigger a sharper second query. Self-checking means the agent inspects whether retrieved evidence supports an answer before generating, and re-retrieves if not. Agentic RAG wins when a question needs more than one lookup or when the right query is not knowable up front. It loses on simple lookups. Each loop turn is a full model round trip, so latency and token cost grow with the number of hops. Use single shot RAG when one retrieval suffices.