Manage AI language models

Install, enable, and disable GEC and Autocomplete models.

AI models power neural grammar correction (GEC) and text autocomplete. You can install them during setup or turn them on/off later.

GEC models: behavior and trade-offs

GEC models generate suggestions from context, not from deterministic rules.

Common trade-offs:

• Output can change after small edits in the sentence.

• Suggestions don’t map to stable “rule IDs”.

  • You can’t reliably ignore “one rule”.

• Dialect nuances are not guaranteed.

  • A single model can be used across dialects.

• Suggestion “types” are imperfect.

  • The UI may classify an AI suggestion as spelling or grammar heuristically.

• AI uses more CPU/RAM and can add latency.

  • Check Hardware requirements.

Try the behavior in a public demo.

Available AI models

Model	Version	Downloaded by default
English GEC	v2.5.0	Yes
English Autocomplete	v1.0.0	Yes
German GEC	v1.0.0	No
Spanish GEC	v3.0.0	No

Install / download models

AI model selection is part of the installation flow. Follow the OS-specific guides for the exact step.

• Linux: Automated installation on Linux (ai_models)

• Windows: Install on Windows with IIS (AI models step)

If the server has no outbound internet during installation, install resources from an offline package.

Use: Manually downloading resources (offline installation)

Verify model files are present

AppServer loads model configs from:

• AppServer/grammar/neural/<lang>/

• AppServer/autocomplete/neural/<lang>/

If a model is missing, you’ll see errors like:

Failed to load config for model 'en'. The model is disabled until the application is restarted. Failed to open AI config file: '<Installation_path>/AppServer/grammar/neural/en/config.json'. Reason: file doesn't exist.

See fixes and more log examples in: Troubleshooting AI model downloads and configuration

Enable or disable AI after installation

There are two controls in AppServerX.xml:

1. <AIModels> controls what models AppServer can load.

2. <Language ...> config controls where models are used.

You should update both:

• Disable AI usage in <Language ...> (check kits + autocomplete).

• Disable the AI model in <AIModels>.

If you break that dependency, you may see: Unknown AI model id 'en'.

Safest way to disable AI (without breaking languages)

1

Back up your configuration

Create a copy of AppServerX.xml before changes.

2

Disable AI usage in language config

For each affected <Language ...> in AppServerX.xml:

• Disable all enabled check kits that include AI.

  • Typical marker: <Neural Enabled="true" .../>

• Make a non-AI check kit the default.

  • Example: lt_hs

To disable autocomplete, disable the <Autocomplete ...> block:

<Autocomplete Enabled="false">
  <Locale>en</Locale>
</Autocomplete>

3

Disable the model in <AIModels>

Set Enabled="false" for every AI model you want to disable.

This step is required. Language config changes alone can still leave AI models enabled and warming up.

If you also want to free disk space, remove model directories after restart and verification.

4

Restart AppServer and check logs

1. Restart AppServer to apply changes in AppServerX.xml.

To restart: stop the application, then start it again.

Use Main AppServer commands.

Verify the model is disabled using status API.

Confirm logs have no AI-related errors.

Example: AIModels section in AppServerX.xml

<AIModels>
	<Model Enabled="true" Lang="en" ModelName="en" Type="grammar" WarmUpId="en_US" SessionsNumber="1" ProbabilityThreshold="0.808" AligningLemmasFilePath="en/en_lemmatization_lookup.txt"/>
	<Model Enabled="false" Lang="de" ModelName="de" Type="grammar" WarmUpId="de_DE" SessionsNumber="1"/>
	<Model Enabled="false" Lang="es" ModelName="de_es" Type="grammar" WarmUpId="es_ES" SessionsNumber="1"/>
	<Model Enabled="true" Lang="en" ModelName="en" Type="autocomplete" WarmUpId="en_US" ProbabilityThreshold="0.15" MaxInputSize="200" SessionsNumber="1"/>
</AIModels>

Disable a model in <AIModels> only after you disabled all usages in <Language ...>. Otherwise AppServer can fail requests with encoder/config errors.