Agentic Retrieval: A Planning-First Architecture for Complex Knowledge Synthesis

ABSTRACT

Traditional Retrieval-Augmented Generation (RAG) models employ a naive 'retrieve-then-read' pipeline, which frequently fails when addressing complex, multi-hop queries that require implicit reasoning. We introduce Cog-RAG, an agentic retrieval framework that embeds a planning and reflection loop prior to database querying. Cog-RAG utilizes an LLM to decompose the user intent, generate sub-queries, and iteratively evaluate retrieved documents for relevance and completeness before synthesizing a final response. Our experiments on the HotpotQA and complex legal datasets demonstrate that Cog-RAG outperforms state-of-the-art dense retrievers by 41% in multi-hop accuracy while completely eliminating out-of-domain hallucinations.