Google Distributed Cloud (GDC) now offers a managed, air‑gapped RAG search service powered by Gemini. The service runs entirely within the customer’s physical security boundary, removes the public‑internet dependency, and provides a turnkey stack for ingesting, indexing, and querying multimodal documents.
Key Changes and Capabilities
The new offering replaces the traditional approach of stitching together separate OCR, speech‑to‑text, vector stores, and custom UI components. Practitioners can upload PDFs, Word, PowerPoint, Excel, plain‑text, HTML, PNG, JPEG, WAV, and MP4 files (up to 100 MB each) and start conversational queries. Gemini returns answers with explicit citations, supports cross‑language queries, and persists conversation state up to 500 MB per chat.
Architectural Overview of the Air‑Gapped RAG Search
All components reside on the organization’s air‑gapped GDC cluster:
- Identity and access: Users authenticate via the existing identity provider; administrators map service roles to provider groups, enforcing per‑user isolation of chats and files.
- Ingestion pipeline: Uploaded files land in an internal object storage bucket. Metadata is recorded in a managed PostgreSQL instance optimized for search. The backend chunks files (e.g., PDF pages or image tiles) and stores the chunks back to the bucket.
- Embedding layer: A Gemini embedding model runs natively in the GDC deployment, converting each chunk into a vector that is stored alongside metadata in PostgreSQL.
- Query processing: When a question is submitted, Gemini rewrites it with conversational context, embeds the revised query with the same model, and performs a vector similarity lookup against the stored embeddings.
- Model serving: The final multimodal prompt—combining the question, history, and retrieved chunks—is sent to a Gemini endpoint executing on NVIDIA B300 GPUs within the same GDC environment. The model streams back a citation‑rich answer.
- Persistence: Each interaction is written to the PostgreSQL database, ensuring that sessions survive across user logins.
Operational and Security Implications
For AI engineers, the managed stack eliminates the need to provision separate OCR, transcription, or vector‑database services, but they must still size the GPU resources to meet latency expectations. Cloud/platform engineers are responsible for provisioning the GDC air‑gapped cluster, configuring the object storage bucket, and maintaining the PostgreSQL instance. DevOps/SRE teams will need to integrate health‑checks for the Gemini endpoint, monitor GPU utilization, and establish backup policies for the persisted conversation data, respecting the 100 MB file and 500 MB chat limits.
Security engineers benefit from the data‑never‑leaves‑premises guarantee. Access control is enforced at the identity provider level, and per‑user isolation prevents cross‑user data leakage. Nevertheless, they should verify that the internal object storage and PostgreSQL are hardened, audit group mappings, and consider network segmentation within the air‑gapped environment to limit blast radius.
Related CloudNinjas coverage: Google Cloud.
What This Means For Practitioners
Adopting the air‑gapped RAG search reduces engineering effort for multimodal enterprise search while preserving strict data‑sovereignty requirements. Teams should evaluate the preview for fit with existing document workflows, plan capacity for the defined file and chat size limits, and align identity‑provider group mappings with internal policies. Ongoing monitoring of GPU health and storage growth will be essential to maintain performance and compliance in a fully isolated deployment.
