Skip to content
[object Object] BOOKSCANNER

Export

Export is the last stage: it assembles the pages you have chosen into a deliverable. The source of every export is each page’s chosen output — the pipeline result selected for that branch — never the raw scan or an intermediate step.

Export targets

TargetWhat you get
PDFa searchable PDF: each page’s chosen image with an invisible OCR text layer on top, so the document looks like the scan but is selectable and indexable
Markdowna free-flowing .md document reconstructed from the OCR — headings, paragraphs, lists, footnotes (from structure where the engine provides it, else inferred from line geometry)
OCR textpackone archive holding all three of the above: the searchable PDF, its Markdown page by page, and the raw response of the OCR engine — see below
Slim projecta pruned copy of the .agl file itself — see The .AGL project file

PDF compression profiles

The PDF exporter picks how each page image is encoded:

ProfileEncoding
autoJBIG2 when every page is bitonal, else keep the originals
g4CCITT Group 4 (universal bitonal fallback)
jbig2JBIG2 — ~25–37 % smaller than G4 (needs the aglaia_jbig2 binding)
nativekeep each image’s original colour/grey encoding verbatim

An OCR text layer is added whenever OCR results exist for the page.

A searchable PDF is searchable, or the export fails

The text layer is invisible, so nobody sees it go wrong. Aglaïa therefore checks it: if a page carries OCR text and no text lands inside the page box, the export fails and deletes the file, and says which pages are missing. An unsearchable PDF that looks finished is worse than no PDF — the library that receives it finds out months later.

A page with no OCR run is owed no text, and is not a failure.

OCR textpack

For a library that wants the scan, the text and the evidence in one file, the OCR textpack (<name>_OCR.textpack) is a zipped TextBundle:

<name>_OCR.textbundle/
├── info.json     what this is, and where the OCR came from
├── text.md       the Markdown, each page preceded by <!-- page N -->
└── assets/
    ├── source.pdf            the searchable PDF
    └── mistral-<job>.jsonl   the OCR engine's raw answer
  • text.md keeps the page boundaries, so a quotation can be cited by page.
  • The raw answer is kept, so the Markdown or the text layer can be rebuilt without paying for the OCR again. Aglaïa stores it in the .agl when the batch result is imported; for a project OCR’d before that, aglaia run <file>.agl --check-ocr fetches it back (a download, not a new OCR).
  • info.json records the provenance: engine, model, job ids, pages, billed pages, cost, and the times.

The format is the one the corpus library reads. With the send-to-corpus plugin installed, the textpack goes straight there.

Markdown refinement

Markdown export can optionally post-process the text with an on-device LLM (--md-refine apple_fm) to repair line breaks and coherence — a no-op when the backend is unavailable.

How to run it

Capture GUIHeadless CLI (aglaia run)
PDFExport tab → PDF--export pdf (or pdf:g4, pdf:jbig2, …)
MarkdownExport tab → Markdown--export md
OCR textpackExport tab → OCR textpack--export textpack (or textpack:g4, textpack:zlib=<id>)
Both—--export pdf:g4+md

Exports are written next to the project as <slug>.pdf / <slug>.md; the textpack takes the original scan’s name, <name>_OCR.textpack.

Each --export entry uses Aglaïa’s standard option-spec format — name[:token|key=value] — so pdf:g4 (token) and pdf:profile=g4 (param) are equivalent, and Markdown refinement is md:refine=apple_fm. The same format drives --ocr (e.g. --ocr apple:lang=fr-FR; use --ocr auto for the default engine); : and = are reserved, quote a value to use them literally.