Skip to content

AI content check

Adaptive Learner can optionally have an AI review a downloaded lesson set (EXP-033). The AI scans the set's cards for translation, grammar and level issues and reports what it finds - it never blocks anything, it only advises. Separately, repositories carry a trust level you can read at a glance.

Repository trust levels

Every lesson set shows a source badge with a trust level in the Content Browser. It speaks to provenance, not to content quality:

  • Trust 0 - not validated. A newly connected repo whose automatic check has not (yet) passed.
  • Trust 1 - technically validated. The repo contains at least one lesson and no executable code. The check re-runs on every sync.
  • Trust 3 - officially recommended. A curated repo from the official recommendation list.

Community ratings (Trust 2) and a central index are not yet implemented.

Prerequisites for the AI check

The AI check calls an AI provider directly from the browser. You need:

  • a stored API key (Settings → Integrations) for one of the providers (Anthropic, OpenAI or Gemini);
  • browser mode (Dexie) - the check runs browser-direct;
  • a downloaded set (the check works on the locally cached cards).

Without a key the button is visible but disabled; a notice links to Settings.

Cost: the check spends tokens on your own provider account and is billed at that provider's rates. Before it starts, the dialog shows a cost estimate you must confirm.

Check a set - step by step

  1. Open the Content Browser and pick a downloaded set.
  2. Click "Check with AI".
  3. The dialog shows a cost estimate. Confirm it to start the run.
  4. The cards are checked in batches; a progress bar tracks it, and you can cancel at any time.
  5. At the end you get a per-card report: only cards with a finding are listed, each with the lesson it belongs to and the AI's note.

There is a short cooldown between two runs (about one minute) so you don't get billed twice by accident.

Apply suggestions (your own sets)

You check your own sets under Content > My content: every set row there carries the same Check with AI button. For a set you created yourself, the report carries the Apply suggestions button. It opens a table with one row per suggestion that names a card field (front, back, notes): lesson, card, field, the current value and the suggested value. Every row is ticked; untick what is an explanation rather than a value. Findings without an applicable value are only counted, never written. The confirm button names the number of fields and cards it changes.

Writing uses the same path as the editor: the whole set with all its lessons, so title, languages, level and description stay, card ids do not change and your learning progress is kept. The previous values are saved and Undo the last apply restores them. After an apply the stored report is out of date and dropped; re-check the set when you want a fresh result. For downloaded sets the button is disabled with a note that it only applies to your own lessons.

Report, cache and the "AI-Checked" badge

  • Cached. The report is stored locally (IndexedDB) and shown again next time without paying again. It carries a content hash
  • a signature, so a changed set suggests a fresh check.
  • Exportable. You can download the report as Markdown - handy for pasting into a lesson revision or an issue.
  • Badge. A checked set shows an "AI-Checked" badge in the Content Browser so you can see a check exists.

The AI is advisory: it highlights possible problems but never prevents you from learning, editing or sharing a set. The call is yours.

For how this works behind the scenes, see the developer docs on AI integration.