The core goal of QueueLingo for WPML is:Manage WPML's translation queue (XLIFF tasks) for automatic translation and automatic delivery back to WPML.You don't need to manually export/import XLIFF, nor do you need to switch back and forth between multiple systems.

In short

WPML generates translation tasks → QueueLingo automatically listens and takes over → Parses XLIFF/HTML translatable nodes → Calls machine translation or LLM → Generates write-back results → Pushes back to WPML → Task completed and records consumption details.

1. Overall Process (from "Task Initiation" to "Automatic Delivery")

Below is a typical end-to-end link:

1. WPML Create Translation Task
   After you create a translation task for an article/page/string in WPML, WPML will add it to the translation queue and generate the corresponding translation task.XLIFF(Or it can be mapped to a structured task of XLIFF).

2. QueueLingo Automatic Listening Queue
   The plugin continuously monitors the task changes in the WPML translation queue, identifies "pending" tasks, and takes over their processing (managed translation).

3. Retrieve tasks and parse content structure
   After QueueLingo receives a task, it parses the XLIFF content structure, locates translatable nodes (segments/units/targets, etc.), and prepares the translation input.

4. Perform translation (MT or LLM)

• Free users: Go for freeMachine translation (MT)Service providers includeGoogle, Edge, Zov

• Subscribers: AvailableLLM Neural Translation(Alternatively, you can enter a custom API Key to bypass the platform's credit limit.)

5. Generate translation results and write them back to the task.
   After translation, the target language text is written back to the corresponding node, keeping the original structure, placeholders, and tags as unchanged as possible.

6. Automatic delivery back to WPML
   Completed tasks will be automatically pushed back to WPML, and once WPML receives them, they can proceed to the subsequent review/publishing process.

7. Record task and consumption details
   The background will generate:

• Translation Task Log(Status, time elapsed, reason for failure, etc.)

• Task Consumption Details(Engine/Model, Token, Billing Method)

2. Why we save on tokens (advantage)

QueueLingo doesn't simply throw the entire webpage/HTML code at the LLM, but instead uses a "Translate only the parts that need to be translated.The strategy:

2.1 XLIFF Task: Accurately Extract Translation Nodes

XLIFF is a structured format designed for localization, with content broken down into many "translation units".
QueueLingo will locate these units from XLIFF, and only...The text that truly needs translationSend it to the translation engine.

2.2 HTML/Rich Text Content: Automatically extract translatable text and replace it with translatable text.

For content containing HTML (such as tags, links, formatted text):

• The system will extractTranslable visible text(Instead of tags, attributes, and structural code)

• Translate the text and then...Precisely replace back to the original structure

This can usually be significantly reduced:

• Invalid token sent to the model (HTML tags, duplicate structures)

• The probability of the model "misclicking a label/damaging the structure"

Result: While ensuring the quality of the translation,The actual token consumption is more economical.The cost is also more controllable.

3. How Translation Engines Work: MT vs LLM

3.1 Machine Translation (MT)

• Suitable for: scenarios requiring rapid process setup, moderate quality requirements, and cost sensitivity.

• Features: Stable, fast, and highly available

3.2 LLM Neural Translation (Subscription Capability)

• Suitable for: Content that is more sensitive to expression, tone, and industry-specific terminology (B2B/B2C/technical documents/marketing creatives, etc.)

• Features: Higher quality ceiling, but cost is related to content length and model price.

4. Prompt Templates and "Website Material Prompts": Making Translations More Business-Oriented

When using an LLM translation website, you can choose the Prompt template:

• b2bSpecialized, formal, and emphasizing value and logic

• b2cNatural, easy to read, and emphasizing experience and emotion.

• tech_docPrecise, consistent terminology, and clear structure

• creativeIt leans towards creative expression, marketing language, and a more flexible style.

It also supports filling in.Custom website profile PromptUsed to supplement:

• Brand tone (e.g., concise and restrained / enthusiastic and lively / authoritative and professional)

• Target audience and product background

• Terminology preferences and disabled expressions

• Translation guidelines for specific fields (such as CTA, button, and title styles)

Glossary feature: Coming soon. Once launched, "terminological consistency" can be further solidified into rules.

5. Execution path with built-in API Key (does not use platform limits)

When you configure the built-in API Key:

• The translation request will use your key to call the corresponding service.

• Does not consume platform credit.

• The platform still provides capabilities such as queue hosting, task tracking, automated delivery, and log and consumption records.

This is perfect for:

• The team already has a model supplier contract/budget.

• I want to consolidate the fees into my own API bill.

• More flexible model selection and quota control are needed.

6. Why does it sometimes fail? How to handle failures and the billing method?

LLM tasks may fail during execution. Common reasons include:

• The model output is unstable (ability fluctuations/“intelligence degradation” phenomenon).

• Timeout, rate limiting, network error

• The input content is complex (placeholders, nested structures, abnormal encoding).

• Invalid key or insufficient balance/quota (more common in scenarios where you have your own key).

  The custom model has insufficient capabilities and limited translation capabilities.

Failure and Deduction Principles

• The system will try its best to retry and tolerate faults (the specific strategy is subject to the product implementation).

• The final charge will be based on actual consumption.:

  • LLM calculates the model unit price based on the actual input/output tokens.

  • Even if the task fails, some costs may still be incurred (e.g., a request has been sent and a partial response has been received), as recorded in the backend consumption details.

7. Where can you watch the whole process?

The backend typically provides two key entry points:

• Translation taskView task status, time elapsed, reason for failure, and retry information.

• Task Consumption DetailsView the engine/model, token, and billing records for each task (platform credit or built-in key is not charged).