Air-Gapped AI: How to Run Generative AI in Isolated and Network-Separated Networks

The short answer: how to run generative AI inside an air-gapped network

If your work network is air-gapped or logically separated from the internet, the way to use generative AI is to run an on-device document AI that processes everything locally on the user's own machine, instead of a cloud service that uploads your documents. Apioni is one such air-gapped AI tool: a desktop app that runs a local language model and a local embedding model on your computer, so your document content stays on your machine, and the only outbound connection it makes is a version check that carries no document content. The point that matters in a regulated environment is not just that it is private, but that your security team can verify the behavior themselves.

This article explains what air-gapped and network-separated environments are, why cloud AI cannot operate inside them, why on-device AI is the structural answer for isolated networks, and exactly how Apioni behaves on the network so you can check it before you trust it.

What "air-gapped" and "network separation" actually mean

An air-gapped network is a network that is physically isolated from the public internet, with no routable path in or out. Network separation, sometimes called network segregation, is the broader practice of keeping a work network logically isolated from an external network, often with two separate environments for internal and internet-facing work.

Both arrangements exist for the same reason: to prevent confidential data from leaving and to prevent external threats from reaching internal systems. They are common in:

• Public sector and government, where internal networks are mandated to be separated from the internet.
• Finance and banking, where customer and transaction data sit behind strict network controls and data residency rules.
• Defense and critical infrastructure, where physical air-gaps are standard.
• Healthcare and research, where patient records and proprietary data cannot be exposed externally.

In all of these, a workstation may have rich internal document access but no usable path to a cloud AI service.

Why cloud AI cannot be used in these environments

Cloud AI tools such as ChatGPT run on external servers and require your data to be sent over the internet to those servers. Inside a separated environment, this fails on three levels at once:

1. The network blocks it. In a true air-gap there is no route to the service at all, so the request simply cannot leave.
2. Policy blocks it. Even where a connection technically exists, security policy commonly prohibits sending internal documents to external services.
3. Compliance blocks it. Uploading internal or regulated documents to a third-party cloud is, in many organizations, a reportable security and compliance violation regardless of intent.

The result is a gap: practitioners know what generative AI can do for document work, but the tools their peers use are structurally out of reach.

Why on-device AI is the structural answer

On-device AI, also called local AI, runs the language model and the embedding model directly on the user's own computer rather than on a remote server. The significance is a reversal of direction: instead of sending the data to the model in the cloud, you bring the model to the data on the machine where the data already lives.

This resolves the conflict cleanly. If the model runs locally and the document content is processed locally, there is no upload to a cloud service, so the network block, the policy block, and the compliance block no longer apply to your document content. The work happens entirely inside the separated environment, which is what makes offline AI viable for network separation.

What the security team will actually ask, and how to answer it

In an air-gapped review, a vendor claim of privacy is not enough. The security team will ask concrete questions, and you need answers they can confirm independently:

• Does any document content leave the machine? The answer must be checkable, not asserted.
• What outbound connections does the app make, and what do they carry? Every connection needs to be accounted for.
• Is there usage telemetry or analytics phoning home? Silent background reporting is a common disqualifier.
• Can we observe this with our own tools? A claim that can only be taken on faith will not pass.

The honest answer to all of these is verifiability. Apioni includes an in-app network self-check, and you can independently confirm its behavior with standard operating system tools such as lsof that list a process's active network connections. In an air-gapped review, "verify it yourself" is what passes.

How Apioni works inside a separated network

Apioni is a Tauri-based desktop app (macOS signed and notarized, and Windows) that runs an on-device LLM, Ollama with a Qwen3-family model, together with the bge-m3 embedding model, locally on your computer. The workflow is straightforward:

1. Connect a folder of documents you want to work with.
2. Apioni indexes them locally, building a searchable index on your machine.
3. Ask questions, summarize, or draft documents, and Apioni answers while citing the source passages it used.

Retrieval uses hybrid search, which combines keyword search (BM25) with semantic search and fuses the two result sets with Reciprocal Rank Fusion (RRF). Keyword search catches exact terms and identifiers, semantic search catches meaning and paraphrase, and fusing them gives more reliable retrieval than either alone. Because answers cite the source passages, you can trace each response back to the underlying document.

Apioni's network behavior, stated precisely

This is the part that matters most in a separated environment, so it is worth stating exactly:

• All document processing goes only to the local engine on your machine (localhost). Your document content never leaves your computer.
• The one connection Apioni makes to the internet is a version check to apioni.com when the app launches. It only checks whether a newer version exists.
• That version check carries no document content, and there is no usage telemetry of any kind.
• You can verify all of this yourself with the in-app network self-check or with standard OS tools such as lsof.

If the machine has no internet connection at all, document processing still works, because that processing never depended on the internet in the first place. Only the version check is skipped, which is why Apioni functions as offline AI inside an isolated network.

Cloud AI vs. on-premise platform vs. Apioni (on-device)

Deploying Apioni on existing workstations

Apioni is built to be deployed by the team that needs it, not by an infrastructure program. It is a one-time purchase of KRW 39,000 with no subscription, a company-wide license with no device limit, and a 14-day free trial that requires no card or payment details. It runs on macOS (signed and notarized) and Windows, on the workstation a person already uses, so a single analyst or department can start without waiting for a procurement cycle or a server build.

You can download Apioni and run the trial inside your own environment first, then buy a license once your security team has confirmed the behavior.

Frequently asked questions

Does Apioni work with no internet connection at all?

Yes. All document processing runs on the local engine on your machine, so it does not depend on an internet connection. With no connectivity, the only thing skipped is the version check to apioni.com at launch; document indexing, retrieval, and answering all continue to work.

Can our security team verify that documents never leave the machine?

Yes. Document processing is directed only to localhost on your machine, and you can confirm this independently. Apioni provides an in-app network self-check, and your team can also use standard OS tools such as lsof to inspect the app's active network connections rather than relying on a vendor claim.

What is the one connection it makes to apioni.com?

When the app launches, Apioni performs a single version check against apioni.com to see whether a newer version exists. This connection carries no document content and sends no usage telemetry. It is the only outbound internet connection the app makes.

Does it need a GPU server or special infrastructure?

No. Apioni is a desktop app that runs on an existing workstation, with the local LLM and embedding model running on that machine. There is no GPU cluster, on-premise server, or integration project required, which is what makes it deployable by a single team rather than an infrastructure program.

Which document types and use cases does it fit?

Apioni is built for people who handle confidential or sensitive documents they cannot upload to the cloud, including legal, accounting and tax, finance, public sector, healthcare and research, and corporate security teams. Typical uses are asking questions across a document set, summarizing long documents, and drafting new documents with cited source passages.

Run it on your own network and verify it yourself

The right way to evaluate AI for an air-gapped or network-separated environment is not to take a privacy promise on faith, but to run the tool inside your own environment and confirm its behavior. Start the 14-day free trial, which needs no card or payment details, install it on a workstation behind your separation, and use the in-app network self-check or lsof to confirm that document content stays local and the only outbound connection is the version check.

When your security team is satisfied, download Apioni for your team or buy a company-wide license. Verify first, then trust.