On-premise and air-gapped document extraction: how it actually works
Air-gapped document extraction means every step, OCR, layout analysis, and the LLM call itself, runs inside your own network with no outbound connection to a vendor's cloud. Not a private endpoint in the vendor's cloud. Not a VPC that still calls out to a hosted model. No egress at all, verifiable by turning off the network and watching the pipeline keep working.
That distinction matters because most vendors selling "air-gapped" mean something narrower.
A regional endpoint is not sovereignty
Picking an EU region on a US-governed platform changes where data is stored, not who can be compelled to hand it over or which law governs the contract. An EU endpoint from a US company is a region toggle, not sovereignty: the company, its governance and its legal framework stay US-based regardless of which region its servers sit in. A platform that is EU-headquartered, an EU legal entity under EU jurisdiction, is a different legal fact than a US platform with an EU checkbox.
Self-hosted is not always air-gapped
Self-hosting your own instance of a vendor's product commonly still means every model call leaves your network to reach that vendor's LLM provider. Ask the specific question before taking "self-hosted" at face value: no outbound network at all, including to the models, or just VPC-isolated compute that still egresses for inference? Most vendors that answer honestly mean the second. A Kubernetes cluster you control that still phones home for every extraction isn't air-gapped, whatever the deployment diagram calls it.
How anyformat's air-gapped deployment works
anyformat's air-gapped mode routes every model call through a local LiteLLM proxy to a local model runtime instead of a cloud LLM provider, and every processing stage, layout segmentation, OCR, rotation, runs on local compute instead of a hosted service. Cloud inference credentials are cleared, not just unused, and telemetry and tracing are disabled for the deployment. The result is a pipeline that produces structured, confidence-scored extraction with zero network calls leaving the perimeter, running the same parse and extract logic as the hosted product.
This isn't a downgraded or stripped-back version of the product for compliance theater. It's the same extraction pipeline, packaged for a customer's own infrastructure as part of the Enterprise tier.
What ships with it
- ISO 27001:2022 certification covering the document processing pipeline end to end, not carved out for a subset of deployment modes.
- GDPR-compliant architecture, with anonymization and deletion tooling built into the product rather than bolted on for one region.
- Zero data retention, available on request. This is a real, configured mode an organization can enable, not the default posture for every account; ask for it explicitly if it's a requirement, and confirm it's active before treating it as a given.
- EU-native governance. The company, not just a data center, is European, which is what actually changes which law applies to a request for your data.
Proof it runs in production, not just in a deployment diagram
The clearest evidence that air-gapped deployment is real and not a slide is a customer running it: the Singapore Prison Service runs anyformat's air-gapped stack in production. That's a government workload, the kind of buyer that verifies a sovereignty claim before signing rather than after, running the local-OCR, local-LLM configuration described above with no cloud calls.
Frequently asked questions
What does "air-gapped" mean for a document extraction platform?
No outbound network connection from the extraction pipeline to any external service, including the LLM provider. OCR, layout analysis and the model call all run on infrastructure inside the customer's perimeter. If a vendor's air-gapped claim still requires an internet connection to reach a hosted model, it's VPC isolation, not an air gap.
Is on-premise deployment the same as air-gapped?
Not necessarily. On-premise usually means the compute runs on infrastructure you control, physically or in your own cloud account. Air-gapped is a stronger claim: no outbound network access at all. A vendor can be on-premise and still call out to a hosted model for every request.
Does an EU region satisfy a data sovereignty requirement?
Selecting an EU region on a platform governed by a non-EU company changes where data is processed, not which country's law applies to a request for that data or the company operating it. A genuinely EU-native, EU-headquartered platform is a different legal position, not a stronger version of the same one.
Does air-gapped deployment mean giving up features?
No, on anyformat it's the same extraction pipeline (schema-based extraction, calibrated confidence, review workflows) packaged for a customer's own infrastructure, not a reduced version built specifically for compliance.
If EU jurisdiction and a verified air-gapped deployment are hard requirements, not a nice-to-have, see how anyformat compares on sovereignty specifically against the vendors evaluated most often alongside it: anyformat vs Reducto, anyformat vs LlamaParse and anyformat vs Unstructured.
