Google details Agentic RAG for cross-corpus enterprise retrieval

Google Research has described an Agentic RAG approach in Gemini Enterprise Agent Platform that is designed to make retrieval-augmented generation more dependable across multiple enterprise data corpora.

What Google announced

In a Google Research Blog post titled “Unlocking dependable responses with Gemini Enterprise Agent Platform’s Agentic RAG,” Google Research says the system addresses a common limitation of conventional retrieval-augmented generation: a single query may not gather enough evidence when the answer depends on information spread across several sources.

The post describes Cross-Corpus Retrieval powered by Agentic RAG as a hosted capability in Gemini Enterprise Agent Platform. According to Google Research, it uses multi-agent query planning, iterative retrieval and “sufficient-context” checks before generating an answer. The goal is to improve whether responses are grounded in retrieved material rather than produced from incomplete context.

Google Research reports that the approach improved factuality accuracy by up to 34% in its evaluation and reached 90.1% on the cross-corpus FramesQA benchmark. Those figures come from Google’s own published research summary and should be read as benchmark results for the evaluated setup, not as a general guarantee for every enterprise dataset.

How the retrieval flow works

Google Cloud’s documentation for “RAG Engine Cross Corpus Retrieval” describes the feature as retrieval across multiple RAG-managed corpora. The documented architecture includes an orchestrator or router, a planning component, a retrieval engine, a reasoning component and a large language model generator.

The same Google Cloud documentation lists two APIs: AsyncRetrieveContexts, which retrieves relevant contexts, and AskContexts, which can generate answers using retrieved material. This distinction matters for enterprise use cases because some teams may want retrieval evidence only, while others may want an answer-generation layer on top of that evidence.

Google’s description suggests that the system can break a user question into subqueries, search across more than one corpus, inspect whether the gathered material is enough and continue retrieving when needed. That design is meant for questions that require synthesis across different document collections, rather than a direct lookup from a single source.

Availability and context

Google Cloud’s Vertex AI release notes state that RAG Cross Corpus Retrieval entered public preview on April 17, 2026. The release notes say the feature can retrieve contexts or generate answers across multiple RAG corpora using AsyncRetrieveContexts and AskContexts.

Public preview means developers can evaluate the feature, but Google Cloud preview offerings may not have the same availability or support commitments as generally available services. Organizations considering the feature should consult the current Google Cloud documentation for region, quota, pricing and production-readiness details.

Why it matters

Retrieval-augmented generation has become a standard pattern for connecting large language models to private or domain-specific information. However, enterprise data is often fragmented across policy documents, tickets, manuals, reports and knowledge bases. A retrieval system that only searches one corpus, or retrieves once without checking whether the evidence is sufficient, can produce incomplete answers.

Google Research’s Agentic RAG work is positioned as a way to make the retrieval step more deliberate. By adding planning and repeated evidence gathering, the system attempts to improve the quality of the context provided to the model before an answer is generated.

The most important claim from Google’s sources is not that the approach eliminates hallucinations, but that it can improve factuality in measured cross-corpus retrieval scenarios. For enterprise teams, the practical question will be how the preview APIs perform on their own corpora, access controls, document quality and evaluation sets.