JSONL output

videre export --jsonl writes one JSON object per file to .videre/hashes.jsonl. It is a way to get the scan inventory into other tools without touching SQLite.

Two outputs, two jobs

They are not competing formats, and neither is the lesser one:

	For
SQLite (.videre/hashes.db)	the library. Everything that accumulates - embeddings, faces, classifications, location clusters - and everything that reads them: search, gallery, dedupe, the MCP server.
JSONL (export --jsonl)	composability. One line per file, straight into jq, awk, a spreadsheet or another program, with no SQLite dependency and no schema to learn.

Use SQLite when videre is your library. Export JSONL when videre is one step in a pipeline you are building. The database is always authoritative; the JSONL snapshot is derived from it on demand.

What JSONL gives up is everything that comes after a scan: it holds the facts about each file and nothing else, so no command reads it back. It is a scan-inventory snapshot, not an annotations or embeddings backup.

videre export --jsonl                       # snapshot every file to .videre/hashes.jsonl
videre export --jsonl --path 2024           # snapshot just one subfolder
videre export --jsonl --dry-run             # count what would be written

videre scan and videre watch never write JSONL; the snapshot is always an explicit export. Each run replaces the previous snapshot atomically, so a scoped export (--path, --type, a date range) writes only the selected files and a selection that matches nothing leaves an empty file, never a partial one.

The format

One JSON object per line, appended:

{"path":"/Photos/IMG_0042.jpg","hash":"5c5254e2...","meta_hash":"9e1f3c07...","size_bytes":4823921,"created_at":"2021-06-14T09:12:33","modified_at":"2021-06-14T09:12:33","ext":"jpg","mime":"image/jpeg","exif_date":"2021-06-14T09:12:33","gps_lat":52.5163,"gps_lon":13.3777,"width":4032,"height":3024}

The fields match the database columns of the same names. Absent values, such as EXIF on a PNG, are null or omitted.

Replaced, not appended. Each export --jsonl writes the current snapshot atomically over the previous one, so the file is always a single consistent picture of the selected files, never an accumulating log.

Working with it

# every HEIC file
jq 'select(.ext == "heic")' .videre/hashes.jsonl

# duplicate hashes
jq -r '.hash' hashes.jsonl | sort | uniq -d

# total size in GB
jq -s 'map(.size_bytes) | add / 1073741824' hashes.jsonl

# photos with GPS, as CSV
jq -r 'select(.gps_lat) | [.path, .gps_lat, .gps_lon] | @csv' hashes.jsonl

# largest ten
jq -s 'sort_by(-.size_bytes) | .[:10] | .[] | "\(.size_bytes)\t\(.path)"' -r hashes.jsonl

Because it is line-delimited, every line-oriented tool reads it directly, so an export feeds straight into a filter:

videre export --jsonl && jq -c 'select(.width > 4000)' .videre/hashes.jsonl

What you give up

JSONL is an export snapshot only. Nothing else reads it:

Command	Works from JSONL?
dedupe, gallery, prune, stats, locations	No, they need the database
embed, faces, classify, search	No
scan (incremental)	No, it needs a database to consult

There is also no perceptual fingerprint, no faces, no embeddings, and no resumability. A JSONL scan is a one-shot description of a folder.

When to use which

Use the database for anything you intend to do with videre itself. It is the default for good reason, and sqlite3 queries against it are usually easier than jq over JSONL. See the database for the schema and example queries.

Use JSONL when videre is one step in someone else’s pipeline: feeding an inventory into another tool, producing an audit log per run, or working on a system where you would rather not keep a database.

If you want structured output from other commands, most support --json, which is a single document rather than a stream:

videre dedupe --json
videre search "sunset" --json
videre stats --json
videre locations --geojson

Those are the better choice for scripting, since they describe results rather than raw scan rows.