Google Cloud, part of Alphabet Inc. (Nasdaq: GOOGL), has previewed Google Distributed Cloud enterprise search with Gemini for air-gapped deployments, giving organizations a way to run retrieval-augmented generation on sensitive material without sending prompts, documents or responses beyond their own physical security perimeter. The service targets sovereign and highly regulated environments where public-cloud AI services cannot handle classified or controlled data.
The managed search capability runs on Google Distributed Cloud air-gapped, the company’s disconnected infrastructure offering. Teams can upload files to a conversational interface and ask Gemini to locate, compare, summarize or translate information, Google Cloud said. Responses cite the source files, and conversations persist between sessions. The company said the preview has no connection to the public internet or Google Cloud’s global infrastructure.
That changes the operating model for organizations that otherwise would have to assemble their own disconnected AI search stack. Google Cloud said that approach can require teams to connect models, speech-to-text and optical character recognition pipelines, vector databases and a user interface, then maintain the resulting environment. The new service packages those functions as a managed capability inside the customer-controlled deployment.
The system accepts PDFs, Microsoft Word, PowerPoint and Excel files, along with text, HTML, image, audio and video formats. Google Cloud said individual files can be up to 100 MB and chats up to 500 MB. After upload, the application stores metadata in a managed PostgreSQL database, divides files into chunks and uses a Gemini embedding model running in the deployment to create vectors for search.
For a query, the company said Gemini first rewrites the question with conversation context, embeds it and searches the vector database. The backend then combines the question, history and matching data into a multimodal prompt for a Gemini endpoint in the same deployment. Because the model can process text, images, audio and video, Google Cloud said it can work across those file types without separate speech-to-text or OCR models.
Keeping these stages local is the central architectural distinction. Google Cloud said uploaded documents, embeddings, prompts and responses remain inside the air-gapped network, and every model runs locally. That boundary is important for IT teams that must meet data-residency, classified-workload or internal security requirements while still giving staff a searchable interface for varied records.
The design also includes local identity integration, role-based access controls and encrypted storage. Google Cloud said data at rest uses AES-256 disk encryption with locally managed hardware security module keys, while data in transit uses TLS with FIPS-approved algorithms. It said the air-gapped platform holds U.S. Department of Defense Impact Level 6 accreditation and is designed to support requirements aligned with FedRAMP High controls.
Google Cloud also said the service can use a local code-execution engine for structured spreadsheet questions. The company said Gemini can generate Python for analytical work inside Agent Substrate, an open-source runtime, while gVisor sandboxes and network isolation constrain that execution. For IT teams, the practical benefit is a supported path to deploy document search and analysis near restricted data, rather than moving that data to a connected AI service or maintaining a custom, disconnected stack.