Unlocking dependable responses with Gemini Enterprise Agent Platform’s Agentic RAG

Details

Google Research and Google Cloud jointly published an agentic RAG framework deployed on the Gemini Enterprise Agent Platform. The system addresses a core weakness of standard RAG: single-step retrieval fails on multi-source, multi-hop queries — for example, finding a server ID in one database and then cross-referencing specs in a separate system. The framework uses a multi-agent architecture with four specialized roles: an Orchestrator that decides when multi-step retrieval is needed, a Planner Agent that maps information pathways across data sources, a Query Rewriter that generates multiple targeted search queries from a single user question, and a Search Fanout Agent that executes parallel searches. A key differentiator is the 'sufficient context' check — the system verifies it has gathered enough information to answer accurately before generating a response, directly targeting hallucination via premature generation. Benchmark results show up to 34% accuracy improvement on factuality datasets compared to standard RAG. Google also reports better grounding and reasoning accuracy on proprietary internal domain-specific datasets. Available now on Gemini Enterprise Agent Platform via RAG Engine and Cross Corpus Retrieval — production deployment, not a research prototype.