Google: Agentic RAG for Gemini Enterprise — 90.1% Accuracy and Up to 34% Better Factuality

Google Research and Google Cloud introduced a multi-agent RAG framework for the Gemini Enterprise Agent Platform on June 5, 2026, whose central innovation is the Sufficient Context Agent — a component that assesses whether the retrieved context is sufficient for a reliable answer. On a demanding benchmark, the framework achieved 90.1% accuracy with significantly improved factuality.

What is RAG and why does it matter?

RAG (retrieval-augmented generation) is a technique by which the model retrieves relevant documents from a knowledge base before composing an answer. Instead of relying solely on what it learned during training, the model grounds its answer in concrete sources.

This yields a double benefit: it reduces the model’s tendency to fabricate facts and increases the factuality of answers. RAG is especially important in business environments where answers must be grounded in a company’s internal documents.

How does the Sufficient Context Agent work?

The biggest step forward compared to classic RAG is the Sufficient Context Agent. Its task is to assess whether the retrieved context is truly sufficient to ground a reliable answer. If the agent concludes that the context is not sufficient, it triggers iterative searches instead of returning an incomplete answer.

This mechanism addresses a common weakness of standard RAG, in which the model answers even when the retrieved material does not fully cover the question. The iterative approach allows the system to gather additional information before forming the final answer.

Which agents make up the architecture?

The framework is organized as a set of specialized agents. There is the Orchestrator, the Planner, the Query Rewriter, the Search Fanout, and the Synthesis agents.

Each agent performs a clearly defined role, and together they form a pipeline that breaks down the query, plans the search, rewrites the query as needed, and finally synthesizes the answer. This division of labor allows for more precise control over each processing step.

What are the results on the FramesQA benchmark?

The framework was tested on the FramesQA benchmark, which covers 824 queries across 2,676 PDF documents. In cross-corpus scenarios, where the answer requires connecting information from multiple sources, it achieved 90.1% accuracy.

Compared to standard RAG, the framework delivered up to 34% better factuality. At the same time, the increase in latency stayed within 3% on average, meaning the gain in quality was achieved without significantly slowing down the response.

What does this mean for business users?

For companies seeking reliable answers grounded in their own documents, the combination of high accuracy and improved factuality with a minimal latency increase represents practical progress. The system is available as a public preview within the Gemini Enterprise Agent Platform, making it available to business users to try out in real conditions.

Particular value is brought by the Sufficient Context Agent, which reduces the risk that the system answers based on insufficient information. In a business context, where a wrong answer can have real consequences, the system’s ability to recognize when there is not enough data and to continue searching rather than guess is just as important as accuracy itself. With this, Agentic RAG targets the reliability that is a prerequisite for serious application in business processes.