Ai Settings

Configure the AI provider, model, and context options used by the Supervertaler for Trados plugin.

Accessing AI settings

Open the plugin Settings dialogue and switch to the AI tab.

Provider selection

Choose one of the supported AI providers:

Provider	Models	Where to get a key
OpenAI	GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, GPT-5.4 Mini	platform.openai.com/api-keys
Claude (Anthropic)	Claude Sonnet 5, Claude Haiku 4.5, Claude Opus 5, Claude Fable 5.1	console.anthropic.com
Gemini (Google)	Gemini 3.8 Flash, Gemini 3.5 Flash-Lite, Gemini 3.1 Pro (Preview)	aistudio.google.com/apikey
Grok (xAI)	Grok 4.3	console.x.ai
Mistral AI	Mistral Large, Mistral Small	console.mistral.ai
DeepSeek	DeepSeek V4 Pro, DeepSeek V4 Flash	platform.deepseek.com
OpenRouter	200+ models from all major providers, one key	openrouter.ai/keys
Ollama (Local)	Runs models on your own machine	No key needed
Custom (OpenAI-compatible)	Any OpenAI-compatible API	From whoever runs the endpoint

Note

You only need one provider to get started. If you would rather not manage several accounts, OpenRouter gives you access to models from all of the above with a single key.

Getting your first key

1. Pick a provider from the table above and follow its link.
2. Create an account and generate an API key. Most providers require billing details before a key will return anything; the free tiers that do exist are usually rate-limited rather than free-forever.
3. Copy the key immediately – most providers show it only once.
4. Paste it into the API key field below, then use Test Connection to confirm it works before you rely on it mid-job.

Keys are stored locally in your Supervertaler data folder and are only ever sent to the provider’s own API endpoint. See Privacy for what does and does not leave your machine.

Caution

An API key is a billing credential. Anyone who has it can spend money on your account. Do not paste one into a shared prompt, a support ticket, or a screenshot – and if you think one has leaked, revoke it from the provider’s console rather than just replacing it here.

API key

Enter the provider’s API key here. Test Connection checks it with a small call.

One key file for every product (from v18.20.187)

Keys are kept in one file shared by Supervertaler for Trados, Supervertaler for memoQ and Supervertaler Sidekick: settings\api-keys.json in your Supervertaler data folder, one line per provider, in plain text. Whatever you type here is written to that file, and the other products read it, so a key pasted once works everywhere. The file is filled from the plugin’s existing keys the first time it is missing, so nothing needs retyping after the update. You can also edit it in Notepad.

The key box says so when a key plainly belongs to another service. Each provider’s keys have a recognisable start: Anthropic sk-ant-, OpenAI sk- or sk-proj-, Gemini AIza, xAI xai-, OpenRouter sk-or-.

Model selection

A dropdown showing a curated list of recommended models for the selected provider.

The short list, and all models (from v18.20.187)

By default the dropdown is a short list: the few models worth recommending for translation work, each with a one-line verdict. It is kept current with every release, and a model that has been superseded is removed from it rather than left in with a warning.

Tick Show all models beside the dropdown to see everything the provider offers as well. Click Fetch list to ask the provider for its current list, using the API key in the box (it need not be saved yet) – Anthropic, OpenAI, Gemini, Mistral, DeepSeek, xAI, OpenRouter, Ollama and custom OpenAI-compatible endpoints all publish one. Fetching switches Show all models on. Models not in the short list appear with the provider’s own name for them; dated snapshots and models that are not for translating text (speech, image, transcription, embeddings) are left out. The fetched list and the tick box are remembered across restarts, and the Batch Operations and chat model menus follow them.

This is the way to use a model released after your plugin build: click Fetch list, pick it, OK.

Model ID

Below the dropdown is an optional Model ID field. To use a model that isn’t in the curated list – a brand-new release, a preview model, or an OpenRouter router such as openrouter/free – type its exact model ID here. When filled, it overrides the dropdown selection; leave it blank to use the model picked from the dropdown.

The field is available for every cloud provider. If you reopen Settings and a saved model isn’t in the curated list, it is shown back in the Model ID field.

Ollama endpoint

When using Ollama as the provider, this field sets the local endpoint URL. Defaults to:

http://localhost:11434

Change this only if you are running Ollama on a different port or a remote machine.

DeepSeek

DeepSeek is a Chinese AI lab offering high-quality models with competitive pricing.

To use DeepSeek directly:

1. Create an account at platform.deepseek.com
2. Go to API Keys and create a key
3. In Supervertaler, select DeepSeek as the provider and paste your key

DeepSeek models are also available via OpenRouter if you prefer a single-key setup.

Custom OpenAI-compatible provider

For providers that expose an OpenAI-compatible API (e.g., Azure OpenAI, together.ai, internal LLM gateways, local inference servers), configure these fields:

Field	Description
Endpoint	The base URL for the API (e.g., https://your-server.com/v1)
Model	The model identifier to use (e.g., llama-3-70b)
API Key	The authentication key for this endpoint

Managing multiple endpoints

You can configure more than one custom endpoint and switch between them without re-entering credentials. Each endpoint is stored as a named profile in the Profile dropdown.

Button	Action
+	Add a new endpoint (starts as “New Endpoint 1”, “New Endpoint 2”, etc.)
−	Remove the currently selected endpoint
✎	Rename the currently selected endpoint

Names are free-form labels – use whatever makes sense for your workflow (e.g. Azure Production, Internal gateway – Mistral Large, Local Ollama). Names must be unique within the list. Renaming is a UI-only change: the endpoint URL, model, and API key all stay attached to the same profile.

OpenRouter

OpenRouter is an API gateway that gives you access to 200+ models from OpenAI, Anthropic, Google, Mistral, Meta, and many others – all through a single API key. Instead of managing separate keys for each provider, you sign up once at OpenRouter and use one key for everything.

Getting started

1. Create a free account at openrouter.ai
2. Go to Keys and create an API key
3. In Supervertaler, select OpenRouter as the provider and paste your key

Curated model list

The model dropdown includes a curated selection of the best models for translation:

Model	Description
Claude Sonnet 4.6	Recommended – best balance of speed, quality, and cost
Claude Opus 4.8	Highest quality – Anthropic’s most capable model, 1M context
GPT-5.5	Premium quality – OpenAI’s most advanced model
GPT-5.4 Mini	Fast, affordable, and high quality for everyday translation
Gemini 3.1 Pro	Google’s most advanced model, large context
Gemini 3 Flash	Fast and affordable – great for large batch jobs
Gemma 4 31B	Open-source – strong multilingual quality, 256K context
Gemma 4 26B MoE	Open-source – near-31B quality at a fraction of the cost
Mistral Small 4	Very fast and cheap – good multilingual support
Qwen 3.6 Plus (Free)	Free – no API costs, good general-purpose quality
DeepSeek V4 Pro	DeepSeek flagship – strong multilingual, competitive pricing
DeepSeek V4 Flash	DeepSeek fast – great for high-volume translation

Using any OpenRouter model

OpenRouter exposes far more models than the curated list above. To use one that isn’t listed, type its exact model ID into the Model ID field (see Model selection) – for example, meta-llama/llama-3.1-70b-instruct, deepseek/deepseek-r1, or a router such as openrouter/free. Browse all available models at openrouter.ai/models.

Pricing

OpenRouter adds a 5.5% platform fee on top of the underlying provider’s token price. For example, if Claude Sonnet 4.6 costs $3/$15 per million tokens at Anthropic, it costs approximately $3.17/$15.83 through OpenRouter. For a typical 5,000-word translation costing $0.50, the OpenRouter fee adds less than 3 cents.

Note

OpenRouter also offers some free models (marked with “Free” in the dropdown). These have no API cost at all – they are rate-limited but perfectly usable for testing or light workloads.

Context layer options

These options control what additional context is included in AI prompts. The settings are split into two groups depending on which features they apply to.

Which settings apply where

Setting	Chat & QuickLauncher	Batch Operations
Termbases in AI prompts	Yes	Yes
Include full document content	Yes	Yes
Max segments	Yes	Yes
Include term definitions and domains	Yes	Yes
Log prompts to Reports	Yes	Yes
Include TM matches	Yes	AutoPrompt only
Surrounding segments	Yes	No

Context layers (Batch operations, Chat and QuickLauncher)

These settings apply to all AI features – Chat, QuickLauncher, Batch Translate, and Batch Proofread.

Include full document content

When enabled, all source segments in the current document are sent to the AI so it can determine the document type (legal, medical, technical, marketing, etc.) and provide context-appropriate assistance. This uses more tokens but greatly improves response quality – the AI can tailor its terminology and style to the specific type of document you are translating.

For very large documents, the content is automatically truncated to the configured maximum. The truncation preserves the beginning and end of the document (first 80% + last 20%).

For Batch Operations, the document content is included once in the system prompt (shared across all batches), so the AI knows what kind of document it is translating even when processing individual batches of segments.

Max segments

The maximum number of source segments to include in the AI prompt when document content is enabled. Default: 500. Range: 100–2000.

Increase this for very large documents where you want the AI to see more content. Decrease it if you want to reduce token usage.

Note

This setting is only available when Include full document content is enabled.

List numbering as structure context (from v18.20.188; always on from v18.20.189)

Word numbers claims, letters steps and bullets lists as paragraph properties, not as text. The segment grid never contains the a) or the 9., and until v18.20.188 neither did anything the AI received. On a patent, that means the model reads six unlettered steps and can conclude the source forgot to letter them – on a real document it translated “steps a. to f.” faithfully and then flagged it as a possible source defect, a note that would have reached the client.

Batch Translate, Translate Segment, Clipboard Mode and SuperBench prefix the first segment of every numbered paragraph with the marker Word renders, inside a sentinel: [#e)]het fixeren…, [#9.]Werkwijze…, [#•]een behuizing…. A rule in the plugin’s own preamble tells the model that this is structure – use it to resolve cross-references and keep list items parallel, never translate it, never include it in the output. The rule ships with every request, including prompts you wrote yourself and AutoPrompt’s, because it lives in the plugin and not in a prompt template. Preview prompt shows exactly what goes out.

The markers come from the original Word file that Studio keeps inside the sdlxliff, and are computed for the whole document at once, so a list that restarts at claim 11 reads 11. and lettered steps that continue from one claim into the next keep counting – exactly as Word shows them. The batch log says how many markers were found and how many segments in the run carry one.

Two safety nets: anything the model echoes back in a target is removed before the segment is written and logged in the batch log, and the TMX backup records the segment as Studio has it, without the sentinel. Files that are not Word documents, and Word files without any lists, get a one-line fallback rule instead: the numbering exists, is not in the text, and its absence is not a defect.

v18.20.188 shipped this behind a checkbox so the first real runs could be checked; they showed no marker reaching a target, so from v18.20.189 it is simply on and the checkbox is gone. If you ever need it off – say, to compare prompts – add "structureContext": false to the aiSettings block of settings.json in %LocalAppData%\Supervertaler.Trados\ while Studio is closed.

Include term definitions and domains

When enabled, term definitions, domains, and usage notes from your termbases are included alongside matched terminology in the AI prompt. This gives the AI deeper understanding of your terminology – for example, knowing that a term belongs to the legal domain or has a specific definition helps the AI use it correctly in both chat responses and batch translations.

Include termbases in AI prompt

Select which termbases are included in AI prompts. Terminology matches from enabled termbases are injected into the prompt to help the AI use the correct, approved terminology.

For AutoPrompt, TermScan automatically filters the termbase to only terms that appear in the document’s source text, keeping the prompt focused and within token limits.

Caution

Only enable termbases you trust. The AI will follow your termbase entries even when they are wrong. If a termbase contains inaccurate, outdated, or low-quality translations, the AI will be forced to use them – producing worse results than if no termbase were enabled at all. Modern LLMs are remarkably good at choosing correct terminology on their own. When in doubt, disable termbases and add terms incrementally as you review the AI’s output.

Context layers (mostly Chat and QuickLauncher)

These settings apply primarily to the Supervertaler chat window and QuickLauncher prompts. The exception is Include TM matches, which also feeds AutoPrompt – see the per-setting notes below.

Include TM matches

The behaviour of this checkbox depends on which feature is asking for context:

• Chat and QuickLauncher (live TM lookups). When enabled, the AI gets translation memory matches – fuzzy and exact – for the active segment. This gives the AI reference translations from your project TMs to improve consistency.
• AutoPrompt (Batch Operations). When enabled, AutoPrompt samples up to 50 already-translated, human-confirmed segment pairs evenly from the active document and includes them in the meta-prompt as in-project reference translations. This includes 100% / exact matches that have been applied and confirmed, fuzzy-and-edited segments, and segments translated from scratch – any segment with a Translated, Approved, or Signed-off confirmation level qualifies. AutoPrompt does not do live TM lookups; it samples confirmed segments straight from the document.
• Other Batch Operations (Translate, Proofread). Unaffected by this checkbox – they always work segment-by-segment without TM reference pairs, regardless of how it’s set.

Tip

Tip for AutoPrompt users: confirm a handful of segments you are happy with before clicking AutoPrompt. Even 10–20 confirmed segments give the AI meaningful style anchors to work from. Without any confirmed segments to sample, the generated prompt won’t have in-project reference translations.

Surrounding segments

The number of segments before and after the active segment to include as context. Default: 5 (five segments on each side). Range: 1–20.

This provides the AI with local context around the segment you are working on. It is also used for the {{SURROUNDING_SEGMENTS}} variable in QuickLauncher prompts.

Note

Batch Operations do not use this setting because each batch already contains a group of segments that provide context for each other.

Tip

Tip: For the best results, enable all context options. The more information the AI has about your project, document, terminology, and previous translations, the more accurate and consistent its suggestions will be.

SuperMemory context

These two toggles control whether SuperMemory content is included in the AI context.

Include memory bank in AI context

When enabled, the AI reads the active bank’s brief.md, terminology.md and style.md - plus the _shared bank of house defaults - before every translation and chat message. This gives it the reasoning behind your decisions, not just the terms themselves.

Caution

Off by default. SuperMemory is a power-user feature that most translators should opt into deliberately. The simpler workflow – TermLens termbases + the AI context options above – covers the majority of needs. Enable this toggle once you have a populated memory bank and want the AI to consult it.

Use memory bank when generating prompts (AutoPrompt)

When enabled, SuperMemory content is included in the AutoPrompt meta-prompt so that generated translation prompts reflect your established client conventions, terminology reasoning, and style guides. Only effective when “Include memory bank in AI context” is also enabled.

Prompt logging

Log prompts and responses to Reports tab

When enabled, AI operations are logged to the Reports tab in the Supervertaler Assistant panel. Each log entry shows:

• The feature and prompt name (e.g. “QuickLauncher · Explain in Context”)
• The model used, estimated token counts, cost, and duration
• Expandable sections for the system prompt, messages, and response

Note

Batch Translate operations appear as a single consolidated entry showing the combined token count, cost, and total duration for the entire operation – regardless of how many sub-batches were processed.

Click “Show system prompt…”, “Show messages…”, or “Show response…” to expand a section. Press Escape to collapse it. Use Copy to copy a single section, or Copy all to copy the full prompt details to your clipboard.

This is useful for:

• Monitoring costs – see exactly how many tokens each operation uses
• Debugging prompts – inspect the full text sent to the AI to understand its behaviour
• Comparing models – run the same prompt with different models and compare token usage

Note

Prompt logging is off by default to keep the Reports tab clean. Enable it when you want to inspect or audit your AI usage. Log entries are stored in memory only and cleared when Trados restarts.

Batch settings

Configure the batch size for the Batch Translate feature. This determines how many segments are sent to the AI provider in a single request.

• A larger batch size is faster but uses more tokens per request
• A smaller batch size is more granular and easier to review

---

See Also

• Prompts
• AI Cost Guide
• TermLens Settings
• Token Usage & Costs