The chatbot answered confidently and got half of it wrong. The model was not the problem.

The search that fed it got exactly one attempt, and nothing in the system was allowed to notice that one attempt was not enough.

That setup is classic RAG (Retrieval-Augmented Generation): search a pile of documents the model was never trained on, say your company's internal wiki or this week's news, paste the best matches into the prompt, and let the model answer from those.

The limit is the shape, not the search. It runs as one fixed pipeline: question in, one search, one answer, out. How many times to search got decided when you built it, not when the question arrived.

Agentic RAG changes one thing. Search stops being a fixed step and becomes a tool, meaning a function the model can choose to call, the same way it might call a calculator or a weather lookup. So the model can read what came back, judge it too thin, rewrite the query, and search again before it commits to an answer.

The picture that made it click for me: classic RAG hands you one photocopied page and says "write your answer." Agentic RAG lets you walk back to the shelves as many times as you need.

Here is the part that bites teams. Handing the agent a search tool is the easy half. Teaching it when the results are good enough is the hard half, and without that judgment it either answers early on thin evidence or keeps re-querying until a hard cap cuts it off.

Quick check before you scroll: A user asks a multi-part question and the first search only answers half of it. What can an agentic RAG system do that a classic RAG pipeline can't?

Full breakdown + the answer: frankduah.me/learnings/2026-09-01-agentic-rag-when-retrieval-is-a-tool-the-agent-calls

New here? I post a bite-size AI / ML concept like this every day - follow me for the daily drop, and it compounds fast. Why I do it: https://lnkd.in/gK8knHDH

#AgenticRAG #AI #LLM #AIAgents #MachineLearning