Retrieval-Augmented Generation (RAG) has become a cornerstone of modern AI systems, but not all RAG implementations are the same. A common interview question is: How is Agentic RAG different from classical single-pass RAG?
Classical Single-Pass RAG works in a straightforward pipeline: retrieve relevant documents once based on a user query, then pass them to an LLM to generate an answer. It's simple and fast, but can struggle with complex, multi-step queries or when the initial retrieval misses critical context.
Agentic RAG introduces an AI agent that can iteratively refine its retrieval and reasoning. Instead of a single retrieval step, the agent may ask clarifying questions, search multiple sources, combine information across steps, and even decide when it has enough data to answer. This makes Agentic RAG more flexible and accurate for complex questions, but at the cost of increased latency and computational overhead.
In short, classical RAG is a linear process, while Agentic RAG treats retrieval as an interactive, goal-driven activity.