The .AGL project file

An Aglaïa project is a single file: <slug>.agl. It is an ordinary SQLite database — no sidecar files, no proprietary format. Everything a project knows lives in it: the raw scans, every pipeline step, the branch choices, the OCR text, the corrections you made by hand on a page, the raw answer of a paid cloud OCR, and the calibration + pipeline snapshots used to produce them.

What it contains

Table	Holds
project	name, slug, schema version, notes (one row)
pipeline_versions	frozen pipeline-YAML snapshots; the active row is current
calibrations	frozen camera-calibration snapshots; the active row is current
images	encoded JPG/PNG blobs (width/height/dpi/type), de-duplicated by sha256
thumbs	per-image thumbnails for the gallery
scans	one row per imported page / captured frame (soft-deleted via deleted_at)
nodes	one row per pipeline-step output; a self-referencing tree per scan
branches	per-branch chosen output (e.g. the A / B halves of a spread)
ocr_runs	OCR results attached to a node
debug_artifacts	optional debug overlays attached to a node
step_overrides	per page layout: skip this pipeline step for this page
manual_overrides	per page layout: the value you set by hand where the pipeline estimated one (rotation, crop, page corners, dewarp)
mistral_batch_jobs	submitted cloud batch jobs and the pages each covers
mistral_batch_outputs	each job’s answer, byte for byte, so the text can be rebuilt without paying for the OCR again

Because images are content-hashed in images, the same pixels are stored once however many nodes point at them, and re-importing a file does not duplicate it.

The slimmed version

A full project keeps every intermediate step (so you can branch, step back, and replay) — which makes it large. A slim copy keeps only what a delivery needs:

• every active scan’s raw root image, and
• every branch’s chosen output (the image an export would use) and its OCR.

All other images, thumbnails, intermediate nodes, debug artifacts, and orphaned OCR runs are dropped, then the file is VACUUMed so its on-disk size reflects the reduced content. The stored raw answers of a cloud OCR are kept: they are what lets the text be rebuilt without paying again. The stored raw answers of a cloud OCR are kept: they are what lets the text be rebuilt without paying again.

Two ways to produce one (GUI Export tab / project menu):

• Export slimmed project — writes a separate slim copy, leaving the working file intact;
• Slim-down current project — rewrites the live file in place (after the project window closes), then reopens it.

A slim project opens normally; it simply has no intermediate stages to step back to.

How to inspect it

It is plain SQLite, so any SQLite tool works — no Aglaïa needed:

sqlite3 my-book.agl '.tables'
sqlite3 my-book.agl 'SELECT id, idx, deleted_at FROM scans;'
sqlite3 my-book.agl 'SELECT id, step_idx, step_name FROM nodes ORDER BY id;'

# pull one stored image out to a file
sqlite3 my-book.agl \
  "SELECT writefile('page.jpg', data) FROM images WHERE id = 1;"

For a GUI, DB Browser for SQLite opens it directly. The images.data column is the raw JPG/PNG bytes.

Related resources

• How Aglaïa works — the file ties all four stages together
• Storage layer — the full schema reference
• Export — including slim-project export