Classic RAG pipelines are inherently fragile because they rely on a strictly linear sequence of steps with no error recovery. If any step fails (e.g., misidentifying a location like Paris, Texas instead of Paris, France), the entire pipeline produces bad results with no mechanism to self-correct. Agentic search addresses this by replacing the linear flow with a loop, allowing the LLM to inspect search results, detect failures, and reformulate queries. The key insight is that optimizing each retrieval step is necessary but not sufficient — you also need an escape hatch for resilience when steps go wrong.
Sort: