Skip to content

Releases: Mkrolick/Anki-Plugin

v0.1.0 — Smart Grader + Book to Cards

26 May 17:48
@Mkrolick Mkrolick
v0.1.0
84b7312

Choose a tag to compare

View all tags
v0.1.0 — Smart Grader + Book to Cards Latest
Latest

First tagged release of the Anki-Plugin repo. Two add-ons, both buildable from source via ./build.sh.

Install

Download the .ankiaddon files below, then in Anki: Tools → Add-ons → Install from file…, pick the file. Restart Anki. Set your OpenAI API key in each add-on's Config.

What's in this release

smart_grader

Typed-answer grading on cards with a Keywords field. On submit:

  1. Strict AND keyword check (case-insensitive, word boundaries).
  2. Cosine similarity between user-answer embedding and reference embedding.
  3. Compared against a per-card calibrated threshold.
  4. Shows pass/fail + word-level diff + similarity number.

Calibration (Tools menu, on selected notes) asks GPT for 12 "good" paraphrases and 6 "bad" near-misses, embeds them all, and fits a threshold so good paraphrases land above and bad ones below. Keywords are AI-generated if the field is empty. The reference embedding is stored on the note so review embeds only the user's answer.

JS injection now happens via mw.reviewer.web.eval() in reviewer_did_show_question (modern Anki strips <script> tags from card HTML).

book_to_cards

PDF → Anki flashcards via a four-stage pipeline.

  • Extract — pdfminer.six (vendored) pulls text per page.
  • Map — sliding 5-page chunks with 1-page overlap; one LLM call per chunk returns {text, quote, topic, source_pages} per noteworthy fact. Source pages are forced to the chunk's actual page list (the LLM's returned indices were unreliable).
  • Merge — embedding-based topic clustering (cosine threshold 0.25), no LLM calls.
  • Reduce — per canonical topic, one LLM call produces 0-5 atomic flashcards. Prompt now includes the full text of every page touched by the topic's aspects as context, plus the verbatim quotes as the actual evidence.
  • Calibrate — runs in a worker thread with ThreadPoolExecutor (default 8 parallel). Embeddings are batched: 19 calls per card collapsed to 1.
  • Insert — main thread, fast: writes pre-computed calibration data to each new note.

The new pre-calibrate stage means Anki's main thread is never blocked during card generation. Resumable across all five stages via JSON checkpoints in the addon's user_files/<book_hash>/.

End-to-end verified on Baudrillard's Simulacra and Simulation (172 pages):

  • 320 atomic flashcards in the deck after Anki's dedup
  • ~16 minutes wall-clock total
  • ~$0.50 in OpenAI charges
  • 4-6× speedup on calibration vs the original serial implementation

Tests

36 unit tests pass; 1 skipped (the deck-writer test needs Anki's bundled Python).

Spec & plan

  • docs/superpowers/specs/2026-05-22-book-to-cards-design.md
  • docs/superpowers/plans/2026-05-22-book-to-cards.md
  • docs/superpowers/reports/2026-05-22-book-to-cards-smoke.md
  • docs/superpowers/reports/2026-05-22-baudrillard-real-book-run.md
Assets 4
Loading