Repository navigation
v0.11.0: Images on cards, gated to concrete nouns
Pictures on cards, for the words that can honestly have one.
A card's word is resolved to a Wikidata concept rather than searched for
as text, which is what makes the judgement language independent: Hund and
chien both reach Q144, so one answer about whether dogs are photographable
serves every language. 45.9% of nouns get a picture, measured over the
whole definition cache: French 43.9%, German 54.3%, English 33.8%.
Off by default. Ask for it per run:
tango run <video-id> --deck "French" --imagesIMAGES_ENABLED=true in .env sets the default for every run, and
--no-images overrides it back for one. A picture roughly doubles a deck
that already carries audio (4.8 MB of 10.1 MB on a real 299-card run), so it
is offered rather than imposed.
Migration
Image and Attribution are appended as fields 12 and 13. tango run
aligns your notetype before importing, so an existing collection gains two
empty fields and keeps every note, its content and its scheduling. Verified
against a live 5,311-note collection, which merged without forking. Anki
will ask for a full sync afterwards, which is what a schema change costs.
What the gate refuses, and why
A wrong picture teaches a wrong association, and unlike a wrong definition it
is not quiet. So the gate refuses abstract concepts by name, after opening
the files rather than reading the rates: leben had returned a photograph of
a newborn, cowardice the Cowardly Lion, government a group portrait of
Dutch ministers, loi the Palais-Bourbon.
It also drops a picture that describes a different sense than the card's
definition. French palais was printing a monumental building beside
"paroi supérieure qui sépare la fosse nasale de la bouche". Four of 214
imaged French words, each read by hand.
Attribution is a licence obligation
Commons reports AttributionRequired: true on the files this returns, so
every picture ships with its credit, and one that cannot be credited is
refused rather than redistributed. Two bugs in that path were found by
counting credits in a real deck rather than by reading the code, and a third
by a pre-release audit: a Commons Artist value is wiki text anyone can
edit, and stripping tags before decoding entities let an escaped payload
become live markup on a card.
Also in this release
- The card is one screen tall, with the picture taking the height the text
leaves rather than a fixed share of the viewport. make distbuilds for PyPI with the README badges pinned, so a release
page stops claiming to be whatever version is newest today.- Ten user-facing messages named a
maketarget that somebody who installed
from PyPI does not have. - A Docker image:
docker pull yousseflarbi/tango:0.11.0.
Full detail in CHANGELOG.md.