Context layers

A translation engine that sees only the sentence in front of it will translate that sentence well and the document badly. Supervertaler’s design is the opposite: every time you send a chat message, translate a batch of segments, or ask AutoPrompt to draft a prompt, it assembles a fresh snapshot of your work and hands the whole thing to the AI.

That snapshot is built in layers. Some come from Trados, some from your own knowledge, and two of them come from parts of the document that no CAT tool normally shows you at all. They stack, and each one that is present removes a class of mistake the AI would otherwise make.

This page is the single place that lists every layer. Each section is a short overview with a link to the feature’s own page.

The ten layers

#	Layer	Where it comes from	On by default
1	Project and file information	Trados	always
2	Current segment	Trados	always
3	Surrounding segments	Trados	always
4	Full document content	Trados	yes
5	A translation memory match	your TMs	yes
6	Termbase terms	your termbases	yes
7	SuperMemory	what you have recorded about the client	yes, when a bank is active
8	List numbering	the Word file inside the sdlxliff	yes, from v18.20.189
9	Figure descriptions	the images in your documents	after two clicks in FigureLens
10	Attached files	you, per chat turn	when you attach something

Layers 1 to 8 need no work from you at all (layer 5 only where Studio has already put a TM hit in the segment). Layer 9 is two clicks per project. Layer 10 is deliberate.

1. Project and file information

Which project and file you are in, the language pair (e.g. Dutch → English), and your position in the document (“Segment 42 of 318”). Always included; no toggle.

2. Current segment

The source text you are translating and any target you have already entered. Always included – the minimum context for most AI operations.

3. Surrounding segments

Two segments before and two after, with their translations where available. This is what lets the AI resolve a pronoun the way the previous sentence resolved it, or continue a clause that began in the segment above. Always included; the window is fixed.

4. Full document content

All source segments in the current document, so the assistant can judge what kind of document it is – legal, medical, technical, marketing, financial, scientific – and let that inform its terminology and register.

Very long documents are truncated to a configured maximum (default 500 segments), keeping the first 80 % and the last 20 % so the beginning and the end both survive.

Toggle: AI Settings → Include full document content.

5. A translation memory match

Where Studio has already put a TM hit in the segment – a pre-translated or auto-propagated row – that translation goes to the AI as work you have already approved, so it is followed rather than quietly rewritten.

Two limits worth knowing, because they are easy to assume away:

It is the match Studio left on the segment, not a search of your TMs for the best few. A segment with no TM origin contributes nothing here, so pre-translating before a batch run is what gives this layer anything to say.
Batch Translate sends exact matches only. Studio records how close a match is but not the source it was made for, so a fuzzy could only be offered as a translation of a sentence the model cannot read – unable to tell which words differ, and most misleading precisely where a memory has been padded with near-misses to manufacture matches. At 100% the match’s source is the segment’s source, so nothing is hidden. Chat, QuickLauncher and AutoPrompt still send whatever match is on the segment, with its percentage.

Toggles: Send 100% TM matches to the AI on the Batch Operations tab, for a job whose memory you do not trust; AI Settings → Include TM matches for the rest.

6. Termbase terms

Matched terms from your active termbases, with approved translations and synonyms, and optionally definitions, domains and usage notes. Non-translatable and forbidden terms are flagged so the AI respects them.

Both Supervertaler termbases and MultiTerm .sdltb termbases attached to the project contribute. See TermLens and MultiTerm Support.

Toggles: AI Settings → Include termbase terms / Include term metadata / the per-termbase contribution list.

7. SuperMemory

SuperMemory is where you record the things about a client that cannot be looked up. If a bank is active, its files go out with every AI call:

brief.md – who the client is and anything standing
terminology.md – term decisions, one table
style.md – prose rules and approved boilerplate

The _shared bank is sent alongside as house defaults, and the active bank is marked as overriding it where the two disagree.

Where a termbase gives the AI flat pairs of terms, SuperMemory gives it the reasoning behind them: the decisions, the caveats, the client-specific overrides. Only the active bank is used (plus _shared), so switch to the right one before translating.

Toggles: AI Settings → Include memory bank context / Use memory bank in AutoPrompt.

8. List numbering

Word numbers claims, letters steps and bullets lists as paragraph properties, not as text – so the segment grid never contains the a) or the 9., and neither did anything the AI received. On a real patent that produced a model reading six unlettered steps, translating “steps a. to f.” faithfully, and then flagging it as a possible defect in the source: a note that would have reached the client.

Batch Translate, Translate Segment, Clipboard Mode and SuperBench now prefix the first segment of each numbered paragraph with the marker Word renders, inside a sentinel – [#e)], [#9.], [#•] – and a rule in Supervertaler’s own preamble tells the model it is structure, to be used for cross-references and parallelism and never reproduced. Anything the model echoes back is stripped before the target is written.

The markers are read from the original Word file Studio keeps inside the sdlxliff and computed for the whole document at once, so a list restarting at claim 11 reads 11. and lettered steps that continue across claims keep counting, exactly as Word shows them. Always on from v18.20.189; see Batch Operations.

9. Figure descriptions

The AI reads your text and cannot see your pictures. FigureLens closes that gap: it takes the images out of the project’s Word documents into a folder, shows each one to the AI together with what the document says about it, and saves a description – what it shows, which figure it is, which parts and reference numbers appear on it – as figures.md in the active memory bank.

From then on it rides along with layer 7 and is read with every request, so the model knows that the “valve (12)” in the sentence is the thing at the top right of Figure 3. Vector drawings (EMF, WMF – what a drawing placed from CAD usually is) are rendered to PNG on the way out, because no AI can read a metafile.

Two clicks per project, then it is automatic. One AI request per image; the button says how many and to which provider before you click, and Describe from the text only is a free alternative that uses what the document itself says about each figure.

10. Attached files

Files you attach to a chat turn – images by paste, drag-drop or browse, and documents (DOCX, PDF, PPTX, XLSX, CSV, TMX, SDLXLIFF, TBX, TXT, Markdown, HTML and more). Images go through each provider’s vision API; documents are text-extracted and appended.

Attachments apply only to the turn you attached them on. See File Attachments.

Stacking the layers

All ten combine freely, and the default composition – everything above except the two that need a click – is a strong baseline and the one to start from.

More context is not automatically better, though. The context window is finite, and a large project with a rich termbase and a mature memory bank can push a prompt into the 50 000–100 000-token range. At some point:

adding TM matches on top of a memory bank that already knows the client’s preferred wordings may add noise rather than signal;
including full document content for a very long document may leave too little room for the memory bank to load;
layering three overlapping sources (TM + termbase + memory bank) on the same concept may produce contradictions the AI has to reconcile on the fly.

The four layers that describe the document rather than the client – 3, 4, 8 and 9 – behave differently: they are cheap, they never contradict each other, and there is no case yet found where turning one off improved a translation.

Seeing and controlling the layers

To see exactly what is being sent, click 👁 Preview prompt next to the Translate button on the Batch Operations tab. It opens a read-only dialog with the assembled prompt – every layer, in the order the model receives it – and makes no AI call.

To change what is sent, open the settings dialog’s AI Settings tab, where you can toggle document content, TM matches, term metadata and memory bank context, and choose which termbases contribute.