sugarcloud.ch · the jexter document engine, over HTTP
Send a PDF or a scan. Get back a book that reads on a phone, clean Markdown for your LLM, a repaired PDF — or the proof that a redaction really hides what it should. One HTTP call, from your server.
No SDK, no upload form: the body of the request is the document. A key per application — ask for yours.
epub-reflowa bookmdfor an LLMdoctagsfor RAGaudita leak foundThe page above, sent to this service. What follows is what came back — not a mock-up.
A reflowable EPUB 3: the text in reading order, cut into chapters with a table of contents, and the chart redrawn as a figure. The reader picks the type size; nothing needs zooming.
Loans by month rose steadily from spring to autumn, as the chart shows. The summer dip is smaller than last year: the garden reading benches kept visitors coming through July.
Three collections grew this year.
Headings stay headings, paragraphs are whole again — no line breaks from the page, no running heads, no page numbers between two halves of a sentence. Ready to chunk and embed.
sample.md
# Lakeside Library — Annual Note 2026 A sample document, written to show what a conversion gives back. # 1. A year of reading The library lent more books this year than in any year before. Most of the growth came from the reading room opened in March, where the new evening ho… Loans by month rose steadily from spring to autumn, as the chart shows. The summer dip is smaller than last year: the garden reading benches kept visi…  Figure 1. Loans per month, 2026.
DocTags, the format of IBM's Docling: each heading, paragraph and picture with its box on the page. An answer built on it can say where it comes from — and point at it.
sample.doctags
<doctag> <section_header_level_1><loc_53><loc_43><loc_382><loc_56>Lakeside Library — Annual Note 2026</section_header_level_1> <text><loc_53><loc_59><loc_304><loc_65>A sample document, written to show what a conversion gives back.</text> <section_header_level_1><loc_53><loc_78><loc_160><loc_86>1. A year of reading</section_header_level_1> <text><loc_53><loc_92><loc_449><loc_118>The library lent more books this year than in any year before. Most of the growt… <text><loc_53><loc_126><loc_448><loc_143>Loans by month rose steadily from spring to autumn, as the chart shows. The sum… <picture><loc_68><loc_163><loc_308><loc_220> </picture> <text><loc_53><loc_229><loc_166><loc_234>Figure 1. Loans per month, 2026.</text>
On page two, an account number sits under a black rectangle. It looks redacted. The audit reads what the page still holds — before the document leaves your hands.
Text the box covers is still text: anyone can copy it out of the PDF.
the audit's answer
{
"pages": 2,
"clean": false,
"findings": [
{
"page": 2,
"kind": "text_under_box",
"box": "v1",
"rect": [149.25, 681.4199, 136.5, 13.5],
"color": "#000000",
"node": "t10",
"text": "CH93 0076 2011 6238 5295 7"
}
]
}Try it yourself: download sample.pdf and send it from the console, or with the calls above.
Sugarcloud is called by your server, never by a web page: the key stays where your code runs.
One key per application, with its quota. Write to contact [at] sugarcube.ch; it is shown to you once.
POST the raw bytes. Its first bytes say what it is: a PDF, an image, or a model you kept.
The answer is the file itself, named after yours — an EPUB, a Markdown file, a PDF, or JSON.
For anything up to a few dozen pages.
# a PDF in, a book out
curl -H "X-API-Key: $KEY" \
--data-binary @report.pdf \
"https://sugarcloud.ch/api/convert?to=epub-reflow" \
-o report.epub
Submit, poll, take the result. A job survives a restart of the service.
# 202 {"uid":"…","state":"QUEUED"} curl -H "X-API-Key: $KEY" --data-binary @book.pdf \ "https://sugarcloud.ch/api/jobs?to=md" curl -H "X-API-Key: $KEY" …/api/jobs/$UID # RUNNING, 42% curl -H "X-API-Key: $KEY" …/api/jobs/$UID/result -o book.md curl -X DELETE -H "X-API-Key: $KEY" …/api/jobs/$UID
Twelve formats, all drawn from one reading of the document — so each is as faithful as the others.
epub-reflowA book: chapters, contents, figures — the text flows to the screen.epubThe pages as they look, fixed layout, with real text to select and search.htmlThe text as one web page: headings and paragraphs.mdClean Markdown in reading order.doctagsDocling DocTags: every block with its box on the page.pdfA repaired PDF: fonts rebuilt, structure regular. A scanned image comes back searchable.svgOne page as a vector drawing.auditIs it really redacted? The text and pictures still under the black boxes, the marks never applied.zonesWhere a pattern occurs — a name, an IBAN — as rectangles to black out.patchOnly what a redaction changed, for a client that holds the model.ocdThe whole model, lossless: send it back later, for any other format.stagesWhat each analysis stage saw — a diagnostic.jexter.ch, the tool site for people, is its first consumer: every tool there is one call here.
One key per application
Shown once when it is made, revoked in a second, never readable back.
Read, never declared
A PDF, an image or a kept model is recognised by its content, whatever the request claims. A scan is read by OCR.
Bounded and durable
A fixed number of conversions run at once; the rest wait their turn. A job a restart interrupted goes back in the queue.
A scan, a two-column paper, a report full of charts. The console takes your key and shows every format side by side.