level: technical
google research and google cloud introduced an agentic retrieval-augmented generation framework for the gemini enterprise agent platform. unlike standard rag, which does a single search and generates an answer, this system uses multiple agents that plan, rewrite queries, and search across different data sources. the key addition is a sufficient context agent that checks if enough information was found before answering. if not, it gives specific feedback to search again, reducing incomplete or guessed responses.
the framework works in phases. an orchestrator breaks down complex queries, a planner maps out needed information paths, and a query rewriter creates multiple search queries. a search fanout agent retrieves snippets from various sources. then the sufficient context agent reviews the retrieved text, an intermediate draft answer, and identifies missing pieces. it can flag gaps like missing allergy data in a patient query and request targeted searches. this loop continues until context is sufficient, after which a synthesis agent produces the final answer.
in tests on the framesqa dataset with 824 multi-hop questions and 2,676 documents, the agentic rag achieved 90.1% accuracy in a cross-corpus setting with four data sources, nearly matching single-corpus performance. this was up to 34% better than vanilla rag. latency remained similar across settings. the system handled questions requiring multiple steps, such as finding tv finale runtimes and calculating differences, by iteratively searching for missing details instead of giving up.
why it matters: this approach makes enterprise ai answers more reliable by verifying context completeness, which is critical for tasks like medical queries or multi-source business analysis where partial answers can lead to errors.