Gemini Enterprise Agent Platform Introduces Agentic RAG: A Multi-Agent System for Reliable Answers to Complex Queries

Researchers from Google Research and Google Cloud have introduced a new multi-agent system for Agentic RAG, deployed on the Gemini Enterprise Agent Platform. Unlike standard RAG, where a single retrieval often yields an incomplete answer for distributed data, Agentic RAG uses multiple specialized agents: a Root Agent decomposes the query, a Planner Agent determines the necessary data domains, a Query Rewriter breaks down long questions into simpler ones, and a RAG Agent performs retrieval. The key innovation is the Sufficient Context Agent, which analyzes retrieved fragments, the draft answer, and the original query, identifies gaps (for example, if an allergy question remains unanswered), and issues specific instructions for re-retrieval. Once all necessary data is collected, a Synthesis Agent formulates the final answer. In experiments on the FramesQA dataset (824 multi-step queries), the system achieved 90.1% accuracy in cross-corpus mode (selecting the correct source from four), nearly matching the accuracy on a single corpus. Compared to Vanilla RAG, accuracy on factual datasets improves by up to 34%. The system is available in public preview on the Gemini Enterprise Agent Platform.