Public sector · Leisure/resort · 案例研究
In an office with no internet, the source passage opens up beside the answer
它能带来什么On a single GPU server inside an air-gapped network with no outside internet, handle summarization, translation, proofreading, reports, and coding from one screen, and check the citation marker and the verbatim internal source passage in the Evidence Panel on every answer.
This is an agency that operates public leisure and resort facilities. Employees work in a network-separated environment, where their PCs and work systems aren't directly connected to the outside internet. Given the nature of public work that handles citizens' information and operational data, security is a non-negotiable premise. The problem was that nearly all the AI tools in use these days are cloud services that run on the outside network. To keep security, you'd have to give up productivity.
So the question they posed was simple. Even in a room with no internet, is an AI that employees can trust and use possible? Consilience built the answer into a demo together with this agency. This article records what that PoC confirmed, and what it hasn't confirmed yet, exactly as it stands.
What's inside the wall, and what's outside
Network separation is, quite literally, splitting the network in two. It severs the work network from the internet — physically or logically — closing off the very path by which data could leak from an employee's PC to an outside website or the cloud. In public agencies it's common, and it's an essential security measure. Here's what this agency has to work with.
- The agency's internal documents. Everything that has to serve as the evidence for an answer sits inside the network
- Employee PCs and work systems. Not directly connected to the outside internet
- One GPU (a high-performance chip for graphics computation) server inside the air-gapped network. Resources are limited, and it has to hold up even when many employees lean on it at once
- Six kinds of work employees want to hand off to AI — summarization, translation, proofreading, report writing, coding, and more
- Recurring tasks a person does by hand, start to finish, every time
Today it's done by hand, or not at all
The wall doesn't just block the bad. Document summarization, translation, proofreading, report drafts, simple coding — all the everyday work people now hand off to AI sits on the other side of it. Employees knew the convenience and still couldn't touch it. You can't break policy. So the recurring tasks stayed a person's job, every time.
And even if you could use an outside AI, another wall remains. In public work, if you can't confirm which document an answer is grounded in, it's hard to trust and use that answer. An answer that can't cite its source — however plausible — can't be put up the approval chain. So this PoC had to handle not only producing answers, but also letting employees confirm where those answers came from.
Not what the AI said, but which line of which document it came from — the user can verify it with their own eyes.
Bring the AI, whole, inside the wall
The direction was clear. Instead of leaving the AI outside and shipping the data out, bring the AI itself, whole, inside the air-gapped network. The model and the data both run on a single GPU server inside the network. On top of that, the agency's internal documents were ingested. Ingesting means organizing scattered internal documents into a search index so the AI can look them up.
How it gets stood up inside the air gap
Load it onto one server as a single bundle
The language model that serves as the AI's brain, the search index, and the app screen the employee faces all go onto a server inside the air-gapped network as one bundle. The whole process of asking a question, finding the relevant documents, and producing an answer finishes inside the network with no outside internet.
Feed in the internal documents with "Extract sources" (⌘⇧I)
The agency's internal documents go into the extraction wizard and come out organized as a search index. Thanks to this prepared index, when an employee asks a question, the system finds the actual internal documents first and uses them as the basis for the answer.
Route work to the model that fits it
Multi-model routing (splitting each task out to a different model) was added. Light summarization or translation goes to a light model; heavy coding or analysis goes to a model suited to it. It was also a choice to make frugal use of limited server resources.
Put concurrent load on it from multiple accounts
The GPU server inside the air-gapped network is limited, yet it has to hold up even when many employees lean on it at once. They actually tested a scenario where multiple accounts placed concurrent multi-turn (several rounds of back-and-forth conversation) load on a single GPU server.
Ask, and check the evidence with your own eyes
"Summarize just the key points of this policy document"
→ A citation marker is attached to the answer. In the Evidence Panel
on one side of the screen, the internal source passage the answer
was based on is laid out verbatim.
"Translate this paragraph and fix anything awkward"
→ Light work goes to a light model. With no need to switch tools,
summarizing, translating, and proofreading all finish on the same screen.
"Draft a report from this"
→ The draft comes with the internal source passages it was grounded in.
Before it goes up the approval chain, the employee reads the original
text themselves and judges.Here's why the answers come out this way. Because the internal documents were ingested in advance and kept as a search index, an incoming question makes the system find the relevant source text first and then build the answer grounded in it. A citation marker is attached to the response, and the Evidence Panel lays that source passage out verbatim. The employee faces one screen. On that one screen, they handle six kinds of work.
Instead of taking the AI's answer at face value, the employee looks at the original text in the evidence panel and judges for themselves. If it's wrong, they catch it immediately; if it's right, they use it. The problem of an answer that couldn't go up the approval chain because you couldn't confirm which document it was grounded in — the evidence panel addresses it head-on.
What changes
| Before | Now |
|---|---|
| The work network is cut off from the outside internet, so summarization, translation, and proofreading can't be handed to AI | On a single GPU server inside the network, six kinds of work — summarization, translation, proofreading, reports, coding, and more — are handled from one app screen |
| Recurring tasks are done by hand, start to finish, every time | With no need to switch tools, the usual work gets done with AI, right where you sit |
| You can't confirm which document an answer is grounded in, so it can't go up the approval chain | A citation marker is attached to the answer, and the Evidence Panel lays out the internal source passage it was grounded in, verbatim |
| Using AI means shipping data out to an outside cloud | The model, the search index, and the app screen run on one server inside the network, so data never crosses the wall |
In sum: inside an air-gapped network with no outside internet, a demo worked in which the model, the search index, and the app screen were loaded onto a single GPU server, six kinds of work were handled from one screen, and every answer came with a citation marker and the internal source text to check. Five passed first-round validation; the complex-document task and the quantitative figures are what's left.