Guide to adopting on-premises (air-gapped) Korean spell checking

In short: To use Korean spell checking in an environment with no external network access, you must install the checking engine on a server inside your organization. Anyone can install the Bareun morphological analysis server and use it. The spell checker uses a lot of resources, so it is provided as a cloud service by default, and air-gapped installation is arranged through a separate consultation.

This guide lists, in order, what to review if you work somewhere that documents cannot leave the premises, such as public institutions, finance, healthcare, and defense.

Why air-gapped environments are a problem

The documents you proofread inevitably include content that must not be taken out, such as contracts, patient records, and internal reports. Because "the moment you paste text into a checker, it leaves the organization," many organizations ban spell checking altogether and rely on people reading by eye.

Pre-adoption checklist

• Scope of outbound data: Check whether the whole text leaves your network or nothing does. With the self-hosted version, documents do not leave the organization.
• Installation specs: You need a CPU with 4 or more cores, 4 GB or more of memory, and 2 GB or more of disk space (System requirements).
• Initial model import: The first start downloads several hundred MB of models and dictionaries. In an air-gapped network, you need a procedure to bring these files in (Model updates).
• Update cycle: Self-hosted installations are updated more slowly than the cloud. Decide in advance when and how to update the models.
• Custom dictionary: Decide who will register your organization's proper nouns and technical terms, and by what procedure.
• Integration method: Decide whether to connect via gRPC, REST, or connect-web, and at which point in your existing document system to insert it.
• License: Morphological analysis is free for research and non-commercial use. Commercial use and the spell checker are paid (License policy).

Choosing an installation method

After installation, the endpoint is <installed host>:5656. Code you used with the cloud service works as is if you only change host (Using a self-hosted installation).

How to bring models into an air-gapped network

graph LR
  A[PC with internet access] -->|Download models and dictionaries| B[Transfer media];
  B -->|Import procedure| C[Air-gapped server];
  C -->|Place in the install path| D[Start Bareun server];
  D -->|Periodic update| B;

If you download the files needed for the first start in advance, the server starts without any outbound connection. The specific file list and where to place the files are described in Model updates.

Using the spell checker in an air-gapped network

Bareun's spell checking adds rule-based correction and an LLM fallback on top of morphological analysis, so it uses far more resources than the morphological analyzer. That is why the default delivery form is the cloud. To bring it into an air-gapped network, you need to decide the following together.

• LLM fallback handling: Either run an LLM inside your organization, or configure the server to use only the rule-based and morphological analysis path without the fallback.
• Hardware: Size the CPU/GPU specs to your throughput.
• Updates: Decide how often to bring in dictionary and model updates.

The answer differs by environment, so please contact technical support and we will work out a configuration with you.

Related documents

• Installation overview · System requirements
• Spell check API comparison — a roundup of checkers you can call from code
• Comparison of enterprise Korean document proofreading services
• Enterprise environments: zero-downtime operation and high availability

Frequently asked questions

Q. Is there a Korean spell checking solution I can use on-premises?

You can install the Bareun morphological analysis server on Docker, Linux, Windows, or macOS. The spell checker uses a lot of resources, so it is provided as a cloud service by default, and air-gapped installation is arranged through a separate consultation. Web checkers such as Naver's have no self-hosted version.

Q. Do documents stay inside an air-gapped network?

With the self-hosted version, correction and analysis finish on a server inside your organization, so documents do not leave. However, if you use an external API for the LLM fallback, text leaves through that path. In an air-gapped setup, you must decide to use an internal LLM or turn the fallback off.

Q. Can it be installed on a server with no internet at all?

Yes. Download the models and dictionaries needed for the first start in advance, bring them in, and place them in the install path. The server then starts without any outbound connection. Later model updates are brought in the same way.

Q. What are the installation specs?

The baseline is a CPU with 4 or more cores, 4 GB or more of memory, and 2 GB or more of disk space. If your throughput is high, add an NVIDIA GPU to increase performance.

Q. Can I use my cloud code as is?

Yes. The self-hosted endpoint is <installed host>:5656 and the API is the same, so you only need to change the connection address in your client.