AI Playground

AI Playground allows you to test your custom AI prompt settings or write a completely custom prompt.

To open the playground:

1. Select AI settings in the project menu
2. Click the OpenAi playground button (or the + Prompt button if some prompts already exist)

You can also quickly get there via the Customize button above machine translation suggestions (only if AI assistant is enabled in suggestions).

Basic mode

In this mode, you can influence what information will be included in the prompt. For example, if you don't want translation to be influenced by Translation Memory results, you can toggle TM Suggestions off.

Advanced mode

info

This feature is only available with the Business subscription plan.

In Advanced mode, you can edit the resulting prompt directly via text input. We use so-called "fragments," which are sections of the default prompt. You can hover any fragment to see its content, and you can also click the Insert button to edit the fragment in the editor.

Other than fragments, you can also use variables, which give you direct access to information about the current key, source and target translations and so on.

Running prompts

You can test your current prompt by hitting the Preview button, and it will run the prompt on the currently selected translation (if the button is disabled, you need to select some translation first).

Once the AI model returns the result, it is displayed in the Result section. This section shows what the resulting translation suggestion will look like. You can also click the </> button to see the raw output of the model.

Preview in batch

You can also run your customized prompt on all currently filtered keys.

1. Click the Arrow next to the Preview button
2. Click Preview all (x keys)

This will let you select which languages to preview and start the batch operation.

After the operation is done, you will see results next to existing translations. The results are only visible to the current user and don't influence currently active translations in any way.

Preview of selected dataset

If you have a bigger project, you might want to have some selected translations that will be used for testing in the AI playground. You can do this by tagging selected keys with the ai-playground tag and then selecting the Preview on my dataset option.

This will run the batch preview only on the tagged keys.

Saving prompts and using them in the platform

To save your prompt, click the Save as... button. Then you'll see the following dialog:

Name your prompt accordingly, and you can also check the Use as default for machine translation if you want to replace the default prompt.

Selecting prompt in language settings

For more granular control, you can go to project language settings:

1. Select Languages in the project menu
2. Go to the Machine translation tab

In this view, you can change each language's machine translation settings globally or individually. Select from your saved custom prompts or reset to the default one.

Handlebars templating system

To use a variable, start by typing {{ (double curly bracket), and the editor will hint you at available variables and fragments.

We use the Handlebars templating system, which has many more functions; read more at Handlebars documentation.

List of available variables

Variables provide access to the translation data so that you can write your prompt as a template and then use it universally for all the translations.

These variables are available in the prompt template:

Name	Description	Example
key.name	Name of the current key	user_settings_save
key.description	User inserted description of the key	A button for saving user settings
project.name	Name of the project	Demo project
project.description	AI customization project description	This is a golfing app
relatedKeys.json	List of related keys based on Tolgee Context	{"source": "Add", "target": "Ajouter", "key": "add-item-add-button"} ...
translationMemory.json	List of related keys based on Translation Memory	{"source": "Add", "target": "Ajouter", "key": "add-item-add-button"} ...
screenshots.first	First screenshot in reduced size (max size 150px x 150px)	[[screenshot_small_*]] (See system placeholders)
screenshots.all	All screenshots in reduced size (max size 150px x 150px)	[[screenshot_small_*]] ... (See system placeholders)
screenshots.firstFull	First screenshot in full size	[[screenshot_full_*]] (See system placeholders)
screenshots.allFull	All screenshots in full size	[[screenshot_full_*]] ... (See system placeholders)
source.*	Source translation/language information	See translation properties
other.{tag}.*	Information for other (not source, not target) translations/languages	See translation properties
target.*	Target translation/language information	See translation properties (non-standard properties are listed below)
target.pluralFormExamples	Lists plural forms for target language with example numerical values (only for plurals)	one (e.g., User has 1 item) other (e.g., User has 10 items)
target.exactForms	Lists plural forms for the target language (only for plurals)	one other
target.exampleIcuPlural	Example format of a plural result for the target language (only for plurals)	{count, plural, one {...} other {...}}

Translation/language properties

Common properties for translation/language of source, target and other.{tag} variables.

Name	Description	Example
.translation	Value of the translation in the given language	Save settings
.languageName	Name of the language	English
.languageTag	Language ISO code	en
.languageNote	AI customization language note	Use remove rather than delete
.isCJK	Is Chinese, Japanese or Korean	false

List of available fragments

Fragments are logical sections for the default prompt. They are accessible the same way as variables, but the difference is that they are using variables within them. They are basically a utility feature that allows you to keep your prompts within a reasonable length.

You can hover any used fragment and click Insert in the tooltip window. This will insert the content of the fragment directly into the editor so you can edit it as you like.

Name	Description
fragment.intro	General info for the model about its role
fragment.styleInfo	Info about resulting translation style
fragment.projectDescription	Provides project.description variable with a description
fragment.languageNote	Provides target.languageNote variable with a description
fragment.icuInfo	Description of ICU format and plurals
fragment.screenshot	Provides screenshots.first variable
fragment.relatedKeys	Provides relatedKeys.json variable with a description
fragment.translationMemory	Provides translationMemory.json variable with a description
fragment.keyName	Provides key.name variable with a description
fragment.keyDescription	Provides key.description variable with a description
fragment.translationInfo	Tells the model what to translate a to what language
fragment.translateJson	Tells the model what the resulting JSON format should be + uses [[output_valid_json]] placeholder, which indicates that the model should return valid JSON (for models that support that).

System placeholders

Rendered prompts can contain system text placeholders, which signify non-text information.

• [[screenshot_*]] signifies a place where the screenshot will be placed
• [[output_valid_json]] signifies that we want to force the model to return JSON. Some models support this, and some don't, so for the non-supporting one, this is just replaced with worded instructions.