Files
step3/docs/FLASHCARDS.md
T
hz 4dfbd811b9 Wire flashcard CSV/JSON import/export and fix theme/responsive gaps
Import and export were stubbed with status-text placeholders; wire them
to the existing backend endpoints with a real file picker (web JS
bridge or native FileDialog) and browser/desktop download, plus an
import preview with create/update/skip/error counts before commit.

Also fix the library and study screens' theme-refresh handlers, which
were no-op stubs so cards never restyled on a theme toggle, and add
responsive layout handling for narrow viewports.
2026-07-21 16:41:26 -05:00

92 lines
5.4 KiB
Markdown

# Flashcards
A separate spaced-repetition system alongside the missed-question review queue. Users author
basic and cloze flashcards (or generate a basic card from a resolved question), organize them
in flat decks, and study them on an FSRS schedule.
## Scheduling
Reviews are scored by [FSRS.Core](https://www.nuget.org/packages/FSRS.Core) 1.0.7, wrapped
behind `IFlashcardScheduler` (`api/QuestionsApi/Services/Flashcards/FlashcardScheduler.cs`) so
the rest of the API never touches the library directly. Pinned, non-configurable defaults:
- Desired retention: 90%
- Maximum interval: 36,500 days
- Interval fuzzing: disabled, so a review's outcome is a deterministic function of stored state
(needed for idempotent replay from the offline outbox, and for golden-vector testing)
- Algorithm/parameter version is recorded per card (`FlashcardSchedule.AlgorithmVersion`) so a
future FSRS upgrade or personalized-parameter feature can be introduced without losing the
ability to tell which cards were scheduled under which version
There is no per-user parameter optimization yet; `FlashcardScheduleStore` stores complete
before/after snapshots on every `FlashcardReviewLog` entry specifically so that a future
optimizer can be trained from full history without a data migration.
`api/QuestionsApi.Tests/FlashcardSchedulerTests.cs` validates the wrapper against the official
FSRS formulas (initial stability/difficulty per rating, and the interval-equals-stability
identity that holds at exactly 90% desired retention) computed from the library's own shipped
default parameters, rather than against hand-copied constants - this stays correct even if
FSRS.Core's defaults are retrained upstream.
## Data model
Six user-owned Mongo collections, all indexed on `(UserId, ...)`:
- `FlashcardDecks` - flat (no nesting); every user has exactly one `IsDefault` deck, created
lazily on first use.
- `FlashcardNotes` - authored content (`Prompt`/`Answer`/`Extra`/`Tags`/`Images`). Cloze notes
store their `{{cN::answer::hint}}` markup in `Prompt`.
- `FlashcardSchedules` - one row per generated card, keyed by `(NoteId, Variant)`. Basic notes
always have variant `0`; cloze notes get one row per distinct cloze number found in the
prompt. Kept in a separate collection from notes so editing a note's content never discards
its review history.
- `FlashcardReviewLogs` - immutable, keyed by the client-generated review UUID for idempotent
submission; carries full before/after schedule snapshots.
- `FlashcardMediaAssets` - user-uploaded images (PNG/JPEG/WebP, 5 MB/4-per-note limits).
- `FlashcardSettings` - one row per user (global new/review daily limits).
Attaching a question's own image to a flashcard does not copy bytes: `FlashcardImageReference`
has a `Kind` (`"question"` vs `"owned"`) so the client can resolve `"question"` references
against the existing `/api/questions/media/{id}` endpoint instead of duplicating the asset.
Sibling cloze cards are buried (`FlashcardSchedule.BuriedUntilUtc`) until the next local study
day whenever any one sibling is reviewed. Daily new/review limits reset at the same boundary.
That boundary is supplied by the client as `studyDayStartUtc` on study/review requests (the
local calendar day at session start, expressed as a UTC instant) rather than computed
server-side, per the spec's client-local-day requirement.
## API
All endpoints live under `/api/flashcards` (`FlashcardsController.cs`) and require `USER`/`SUPER`
auth. Deck/note mutations use the same optimistic-concurrency (`ExpectedRevision` + 409 Conflict)
pattern as the rest of the API. Review submission is idempotent on `ClientReviewId`; undo only
allows reverting the single most recent review on a card (it refuses if the card's revision has
moved since).
## Import/export
CSV is UTF-8, text-only, headers `stable_id,deck,type,prompt,answer,extra,tags,
source_question_external_id` (tags semicolon-separated); it omits images, schedule state, and
review history and reports those omissions via the `X-Flashcard-Export-Omissions` response
header. App JSON is full-fidelity (decks, settings, notes, schedules, review logs, base64
images) and merges on import by the note's stable UUID. Both formats go through a
create/update/skip/error preview (`FlashcardImportPreviews`, TTL-expired after an hour) before
`import/commit` applies it; commit never silently drops a row - everything is created, updated,
skipped as an exact-content duplicate, or reported as a row-level error. Limits: 25 MB and
10,000 notes per import.
The Godot client (`scripts/autoload/flashcard_service.gd`) picks a file, uploads it to the
matching preview endpoint as `multipart/form-data`, and shows the create/update/skip/error counts
before letting the user commit or cancel. Export streams the response body and saves it client
side. File selection and download both branch on `OS.has_feature("web")`: on web, a hidden
`<input type="file">`/`FileReader` pair (bridged back into GDScript via
`JavaScriptBridge.create_callback`) and an anchor-click data-URL download stand in for a native
picker/save dialog; on desktop, a `FileDialog` and the OS downloads folder are used instead.
## Deployment
No new secrets. `FlashcardsController`/stores register the same way as the rest of the API
(`Program.cs`, `AddSingleton`). The Godot client resolves its API base URL from
`step3/flashcards_api_url` in `project.godot`, following the same same-origin-on-web resolution
as every other autoload service (`api_endpoint_config.gd`).