Gemini enterprise search on Google Distributed Cloud (GDC) air-gapped is now in preview. It is a fully managed retrieval-augmented generation (RAG) search experience that runs entirely inside an organization’s physical security perimeter, with no connection to the public internet or to Google Cloud’s global infrastructure. The target is national security agencies and other sovereign or highly regulated organizations whose data cannot leave customer premises. The earlier alternative was custom agentic search inside a disconnected perimeter, which meant months of stitching together models, building speech-to-text and OCR pipelines, configuring vector databases, and maintaining a bespoke UI.

Users upload files into a conversation and ask Gemini to find information, summarize it, compare it, or translate it. Every answer cites its sources, and conversations persist between sessions. The service also supports cross-language questions, and users can ask one question across many formats — audio, video, images, and spreadsheets — without separate transcription or OCR steps. Each user’s chats and files are private to that user, and access follows the identity provider groups administrators already manage. Supported file types include PDF, Word, PowerPoint, Excel, TXT, HTML, PNG, JPEG, WAV, and MP4, with each file up to 100 MB and each chat up to 500 MB.

Under the hood, a query moves through several layers, all on the organization’s air-gapped network. Users sign in through their organization’s identity provider, and administrators map the service role to identity provider groups. When a file is uploaded, it goes to the app’s object storage bucket, and the backend records metadata in a managed PostgreSQL database optimized for search. The backend splits the file into chunks such as groups of PDF pages or tiles of an image, then a Gemini embedding model running natively within GDC converts each chunk to a vector stored in PostgreSQL. For a question, a state-of-the-art Gemini model running entirely on GDC first rewrites the question with conversation context, then the backend embeds the rewritten question with the same model and searches the vector database.

Model serving combines the question, conversation history, and best-matching data into a single multimodal prompt and sends it to a Gemini endpoint running on NVIDIA B300 GPUs in the same GDC deployment. Because Gemini is natively multimodal, it can accept text, images, audio, and video as input and reason across file types in one context, so no separate speech-to-text or OCR models are needed. Answers stream back to the user with citations to source files. The persistence layer stores each question and answer in the system database and keeps uploaded files in object storage, allowing users to reopen an earlier session and continue.

For classified and sovereign environments, prompts, responses, uploaded documents, and embeddings never leave the air-gapped network, and every model runs locally. 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 with granular role-based access control governs resource access. GDC air-gapped has U.S. DoD IL6 accreditation and is designed to support strict requirements aligned with FedRAMP High security controls.

For deterministic analytical answers from structured spreadsheet data like CSV and XLSX files, the service uses the Gemini inference stack’s local code-execution engine. Gemini dynamically generates Python code and runs it securely within Agent Substrate, an open-source execution runtime built for large-scale agent orchestration. Agent Substrate enforces strict network isolation and runs workloads inside gVisor sandboxes, mitigating the security risks of executing LLM-generated code without sacrificing inference performance.

The announcement positions this as ending the tradeoff between security and frontier 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.