Managed RAG search inside the sovereign perimeter: Gemini on GDC

The first draft ran long. I’m tightening each section to the required paragraph length while keeping the names, limits, and architecture details.Google announced enterprise search with Gemini on Google Distributed Cloud (GDC) air-gapped on October 20, 2026, in Google Cloud Blog's report by Rohan Grover, Senior Director, Google Distributed Cloud. The preview is fully managed retrieval-augmented generation (RAG) search inside the physical security perimeter, with no connection to the public internet. Teams upload files and ask Gemini to find, summarize, compare, or translate information. Every answer cites its sources, and conversations persist between sessions.

This piece is for national security agencies and other sovereign and highly regulated organizations. They keep classified reports, field manuals, and sensitive multi-modal datasets on customer premises under strict data sovereignty mandates. It covers the managed preview that replaces a custom stack, how a query is answered on premises, the file caps, and the encryption, U.S. DoD IL6, and code-execution controls.

Gemini on GDC: what actually changed

Until this preview, the alternative was custom agentic search inside a secure, disconnected perimeter. Grover writes that the build took months: stitching models together, building speech-to-text and optical character recognition (OCR) pipelines, configuring vector databases, and maintaining a bespoke user interface. Enterprise search with Gemini on GDC air-gapped replaces that custom stack and has no connection to the public internet. Prompts, responses, uploaded documents, and embeddings never leave the air-gapped network, and every model runs locally inside the data center.

Users can go from hundreds of pages to a cited answer across manuals, reports, and policies, and see which files each answer came from. They can ask in one language and get answers grounded in documents written in another, or have Gemini translate them. One question can combine audio recordings, video clips, images, and spreadsheets with no separate transcription or OCR step. Chats and files persist, so work does not start from scratch. Need-to-know access follows identity provider groups, and chats and files stay private.

Gemini on GDC: how it works

A query crosses four layers on the air-gapped network: signing in, ingesting files, generating the Gemini response, and storing the session. Users sign in through their organization's identity provider. Administrators map the service's user role to identity provider groups, and only those members can access the application. A file goes to the app's object storage bucket. The backend records metadata in managed PostgreSQL optimized for search. It splits the file into chunks, such as groups of PDF pages or image tiles, and writes the chunks to the bucket.

A Gemini embedding model running natively on GDC turns each chunk into a vector stored in PostgreSQL. The state-of-the-art Gemini model in the customer's data center rewrites the question with conversation context. The backend embeds that rewrite with the same model and searches the vectors. A multimodal prompt of the question, the history, and the best-matching data goes to a Gemini endpoint on NVIDIA B300 GPUs. Gemini accepts text, images, audio, and video without separate speech-to-text or OCR models. The cited answer streams back.

Gemini on GDC: why it matters now

Google says organizations no longer have to choose between uncompromising security and cutting-edge multimodal intelligence. Enterprise search, Gemini 3.5 Flash, and databases optimized for search run together inside the GDC air-gapped perimeter with no dependency on the internet. Google describes that bundle as the way to bring the best autonomous capabilities to the sovereign cloud. Find, summarize, compare, and translate run in one cited conversation, with that stack running in the customer's data center rather than on Google Cloud's global infrastructure.

Data is encrypted at rest with AES-256 disk encryption using locally managed HSM keys, and in transit with TLS using FIPS-approved algorithms. Local IAM integration with granular role-based access control (RBAC) governs access to resources. GDC air-gapped has U.S. DoD IL6 accreditation and is designed to support strict requirements in alignment with FedRAMP High security controls. Embeddings never leave that network: the local Gemini embedding model writes them into the deployment's PostgreSQL database, while source files stay in object storage.

Gemini on GDC: who is affected

The customers named are national security agencies and other sovereign and highly regulated organizations. Their staff members work with classified reports, field manuals, and sensitive multi-modal datasets that cannot leave the premises. Also affected are teams that maintained the custom path: custom models, speech-to-text, OCR pipelines, vector databases, and a bespoke user interface. Only members of identity provider groups mapped to the service user role can open the application, and each user's chats and files stay private to that user.

Supported files are PDF, Microsoft Word, PowerPoint, and Excel, plus TXT, HTML, PNG and JPEG images, WAV audio, and MP4 video. Each file can be up to 100 MB, and each chat up to 500 MB. For deterministic answers from CSV and XLSX data, such as the average salary across departments, the service uses the Gemini inference stack's local code-execution engine. Gemini dynamically generates Python and runs it in Agent Substrate, an open-source execution runtime for large-scale agent orchestration, inside gVisor sandboxes with strict network isolation.

Gemini on GDC: what to watch

The service is in preview. Google points organizations to the GDC air-gapped documentation and to their GDC representative for sovereign-cloud use. Caps already published are 100 MB per file and 500 MB per chat, covering PDF, Microsoft Word, PowerPoint, Excel, TXT, HTML, PNG, JPEG, WAV, and MP4. Only mapped identity provider groups can sign in. Questions, answers, and files stay in the local database and object storage, with no connection to the public internet or to Google Cloud's global infrastructure.

CSV and XLSX questions are answered with Python that Gemini generates and Agent Substrate runs. That open-source runtime enforces strict network isolation and uses gVisor sandboxes against the risk of LLM-generated code, without sacrificing inference performance. The Gemini embedding model, the state-of-the-art Gemini model that rewrites the question, and a Gemini endpoint on NVIDIA B300 GPUs all run in the GDC deployment. Gemini 3.5 Flash is named with enterprise search and the search-optimized databases. Prompts, responses, documents, and embeddings stay on the air-gapped network.