Agentic RAG in Practice: Multi-Step Research Workflows

Tracing one question through the loop

Lesson 17 introduced agentic RAG's three capabilities in the abstract. Take the multi-hop question from Lesson 5, "which of our vendors raised prices after switching CEOs," and trace it through an agentic RAG loop step by step, to see how planning, tool use, and iteration actually combine in practice.

Step by step

Plan: the agent recognizes this cannot be answered by one retrieval, and breaks it into two sub-questions: "which vendors switched CEOs recently" and, for each result, "did that vendor also raise prices." Retrieve (sub-question 1): the agent chooses a document search over vendor news and retrieves a list of vendors with recent CEO changes. Evaluate: the agent checks whether this result is specific enough to proceed, here it is. Retrieve (sub-question 2, iterated per vendor): for each vendor found, the agent runs a separate, targeted retrieval for pricing announcements, rather than one generic query across the whole vendor list. Synthesize: only after all sub-question results are in does the agent generate the final answer, cross-referencing the two retrieved sets.

Why this could not have been one retrieval call

No single embedding of the original question would have found this answer, because the answer is not a fact stated anywhere, it is the intersection of two separately-retrieved facts. This is the same structural gap graph RAG (Lesson 12) closes through explicit relationship traversal; agentic RAG closes a version of the same gap through explicit multi-step retrieval planning instead, without requiring a pre-built knowledge graph. The two are not mutually exclusive: an agentic system can use a graph-query tool as one of the tools available to it during planning.