On-premises document extraction: how to evaluate it

What is on-premises document extraction?

On-premises document extraction means the software that reads your documents runs on servers you control. The term gets stretched, so ask where three things happen: where pages are read, where results are stored, and what still crosses your boundary. A product can process files locally and still send usage data out for billing. Deployment options sit on a line, from least to most isolated.

What does air-gapped actually mean?

An air gap is a precise term. The NIST glossary defines it as an interface between two systems that are not physically connected, where any logical connection is not automated, so data crosses only manually and under human control. Blocking most outbound traffic is therefore not an air gap, and neither is running on your own hardware. Microsoft’s documentation shows the difference. Its connected containers run locally yet send usage information to the cloud for billing, while disconnected containers are intended to run with no cloud connectivity. Quiet connections are easy to miss, too. The widely used huggingface_hub library collects some telemetry by default and offers a setting to turn it off, along with an offline mode. Any stack built on open components needs that level of checking.

When do you need on-premises extraction?

You need it when a rule, a contract or the physical site prevents documents from leaving. Microsoft’s eligibility criteria are a fair summary of the real cases: environments with no internet connection, remote locations with occasional access, and organisations under strict regulation against sending any data to the cloud.

Be honest about the cost. You take on hardware, patching, monitoring and capacity planning, and a new model reaches you only when someone carries it in. If your concern is simply that documents stay in an account you control, a private VPC may meet it with far less work. Ask your security team which boundary the requirement names before you buy for the strictest one.

What on-premises options exist today?

Four kinds of option exist, and they differ more in access and scope than in the label on the box.

The details matter. Microsoft says access is limited to customers identified as strategic customers or partners with a qualifying use case, reviews requests within 10 business days, and charges the full price of a calendar-year commitment plan immediately. Its containers also need a licence file with an expiry date. Reducto describes offline update bundles for on-premises and air-gapped sites, verified with vendor checksums. The Docling paper describes fully local execution suited to sensitive data and air-gapped environments, on commodity hardware. Docling converts documents; turning that output into named, checked fields for accounts payable is left to you.

How should you evaluate an on-premises product?

Evaluate the offline product, not the cloud demo. The cloud version and the installed version can differ, and the gaps tend to appear in the parts that matter after go-live. Check these six areas.

• Feature parity. List what the installed version lacks compared with the cloud one, including review screens, custom document types and new models.
• Accuracy parity. A smaller model that fits your hardware may read worse than the hosted one. Test on your own documents, not a published benchmark.
• Network behaviour. Confirm what the software tries to reach, by watching traffic rather than reading a datasheet.
• Updates. Find out how models, patches and licences enter a closed network, how often, and how you verify them.
• Capacity. Get hardware figures for your volume and your longest files. As one reference point, Microsoft recommends 8 cores and 24 GB of memory per Document Intelligence container.
• Operations. Decide who is called at night, what logs exist, and how you diagnose a failure when the vendor cannot see the system.

What should you ask every vendor?

Ask these in writing, and treat a vague answer as a no until it is shown working.

1. Does anything leave the network in normal operation, including licence checks, usage reports, telemetry and crash logs?
2. Does the software stop if a licence expires, and how is a licence renewed with no connection?
3. Which features are missing from the air-gapped version?
4. Do the models run entirely inside our network? Does any step call an outside model?
5. How are updates delivered, how often, and how do we verify them?
6. Can new document types be added offline, and who does that work?
7. Are corrections kept on our side, and does any of our data train a shared model?
8. What hardware do we need for our volume, and do you require GPUs?
9. How does it handle long, multi-page files and bundles of mixed documents?
10. Who has access during support, and how does support work with no remote session?
11. Which certifications apply to the installed product rather than to your cloud service?

The last question catches a common confusion. A cloud certification describes the vendor’s hosted environment. Software you install runs on your controls, so a certificate does not automatically follow it, and self-hosting does not by itself make a product compliant.

How do you run a proof of concept?

Run the proof of concept in the same conditions as production, because a test on an open network proves little.

1. Install on an isolated network with outbound traffic blocked and logged.
2. Process a sample of real documents, including poor scans and your longest files.
3. Measure accuracy per field against values a person has checked, and note which errors were flagged and which were silent.
4. Review the outbound log. Each blocked attempt needs an explanation.
5. Rehearse an update by carrying in a patch or new model by hand.
6. Pull the licence or stop a service, and see how the failure appears.

Where does docRefinery fit?

docRefinery is a document data extraction platform that runs in managed cloud by default. On the Enterprise plan it runs in your private VPC or fully air-gapped on premises, on models that run inside your network, so a document never has to leave it. It suits back-office work where the same senders repeat: each layout is learned once and remembered, you define every field’s instruction and output key, and every document type returns one format. Long, multi-page files are segmented into the documents inside them. Extracted values carry a confidence score and a page citation by default, an absent field comes back null, low-confidence values go to a review queue, and corrections become suggested fixes.

Hold us to the same list. This is a deployment capability, not a certification: our SOC 2 Type II audit is in progress, with no report yet, and SSO is planned for Enterprise. If you want documents in a vector store, a retrieval pipeline is planned but not available, and a parser built for that job will serve you better today. Contact us to scope a deployment, or see what each plan includes.