🧠 Anki Language Sentence Mining
AI-powered sentence generation for Anki — Turn your vocabulary cards into natural, contextual sentences using Google’s Gemini AI.
Sentence Miner bridges the gap between memorizing isolated words and understanding them in real conversation. It reads your vocabulary deck, identifies which words you know well and which you’re still learning, then generates sentences that mix them together — just like how you’d encounter them in the real world.
✨ Features Two Card Types 🔲 Cloze Recall: Hides known words so you recall them from context. (Example: 猫が{{c1::屋根}}の上に座っている) 🆕 New Word: Introduces unfamiliar words alongside known ones. (Example: 猫が屋根の上に座っている)
You control the mix with a slider (default: 70% cloze / 30% new word).
Smart Word Selection
The add-on doesn’t just randomly pick words. It uses several strategies to create better sentences:
Tag-Aware Sampling — If your vocabulary cards are tagged (e.g., food, travel, JLPT_N3), the generator groups words by topic and creates thematically coherent sentences instead of random word combinations. Priority-Weighted New Words — Words you struggle with (high lapse count, low ease factor) or haven’t seen yet are prioritized for introduction. This means the sentences focus on words that need the most reinforcement. Confidence Threshold — You define what “known” means by setting a minimum review interval (e.g., 7 days). Cards above that interval are considered known; everything below is a candidate for introduction. Performance Optimized In-Memory Cache — Word lists are cached for 5 minutes, so hitting “Generate” multiple times doesn’t re-query your entire deck each time. Note-Level Deduplication — Fetches each note only once even if it has multiple cards (e.g., forward + reverse cards), preventing duplicate words. Background Generation — API calls run in a background thread so Anki’s UI stays responsive. Natural Sentences
The Gemini prompt is carefully crafted to produce sentences that:
Sound like something a native speaker would actually say Fit daily conversation contexts Match the learner’s current level Are NOT awkward combinations of random vocabulary Natural Grammar & Conjugation Context-Aware Conjugation — The AI understands that words in your deck are dictionary forms (lemmas). It automatically conjugates verbs and adjusts nouns/adjectives to fit the sentence context naturally (e.g., changing “go” to “went”). Smart Cloze Matching — Even when a word is conjugated, the cloze deletion handles it correctly, hiding the actual used form while preserving the underlying concept. Optional Grammar Breakdown — Get a word-by-word analysis of grammar, particles, and composition added to your card notes. 📦 Installation
Install from GitHub or AnkiWeb with code.
- Restart Anki
Close and reopen Anki. You should see a new Sentence Miner submenu under Tools.
🔑 Setting Up Your Gemini API Key (Free)
On first launch, the add-on will prompt you to enter an API key. You can also access this anytime via Tools → Sentence Miner → API Key Setup.
Go to Google AI Studio — API Keys Sign in with your Google account Click “Create API Key” Select any Google Cloud project (or create one — it’s free, no credit card required) Copy the key and paste it into the add-on dialog Click “Test Key” to verify it works Click “Save” Free Tier Limits gemini-2.5-flash: 10 Requests/Min, 250 Requests/Day gemini-2.5-flash-lite: 15 Requests/Min, 1,000 Requests/Day
This is more than enough — generating 5 sentences uses 1 request.
🚀 Usage Generating Sentences Go to Tools → Sentence Miner → Generate Sentences Configure the settings: Source Deck: The vocabulary deck to pull words from Language I’m Learning: Target language (e.g., Japanese, Spanish) My Native Language: Language for translations (e.g., English) Known Word Threshold: Minimum interval (days) for a word to count as “known” Known Words in Sentence: Target percentage of known words per sentence Card Type Mix: Ratio of Cloze cards vs New Word cards Sentences to Generate: How many sentences per batch (1–20) Sentence Deck: Name of the deck where generated cards will go Include Grammar Breakdown: (Checkbox) Add word-by-word analysis to notes Click ⚡ Generate Sentences Review the generated sentences in the preview table Uncheck any sentences you don’t want Click ✅ Add Selected to Deck Settings Are Remembered
All settings are saved automatically and restored next time you open the dialog.
📋 Card Structure AI Sentence (Basic) Sentence: The target-language sentence with new words in bold Translation: Natural translation in your native language NewWords: List of new words with readings and definitions Notes: Additional context (if any)
Front shows the sentence with a hint button for new words. Back reveals the translation and word definitions.
AI Sentence Cloze Text: The sentence with {{c1::hidden_word}} cloze deletions Translation: Natural translation in your native language Notes: Additional context (if any)
Front shows the sentence with blanks. Back reveals the hidden words and translation.
Both card types include styled templates with:
Clean typography (Noto Sans family) Dark mode support Color-coded new words (orange) and cloze words (blue) ⚙️ Configuration
You can edit the config via Tools → Add-ons → Sentence Miner → Config
🧩 How It Works Under the Hood Word Classification Your Vocabulary Deck │ ├── Cards with interval >= threshold ──→ KNOWN WORDS │ (e.g., interval >= 7 days) Used for cloze blanks │ and sentence scaffolding │ └── Cards with interval < threshold ──→ NEW/WEAK WORDS (new, learning, or young cards) Introduced in context Sorted by priority: with known-word support • High lapses = higher priority • Low ease = higher priority • Brand new = high priority
Smart Sampling Pipeline All Known Words (e.g., 3,000) │ ├── Group by Anki tags │ food: [rice, fish, cook, …] │ travel: [hotel, ticket, map, …] │ JLPT_N3: [決める, 届ける, …] │ ├── Pick 1-2 focus tag groups │ ├── Sample 70% from focus groups ──→ Thematic coherence └── Sample 30% from others ────────→ Variety │ 50 words to prompt
All New Words (e.g., 500) │ ├── Calculate priority score per word │ score = lapses×2 + (3-ease) + new_bonus │ ├── Sort by priority (with jitter) │ └── Sample from top candidates ──→ 20 words to prompt
Generation Flow [Smart-sampled words] ──→ [Gemini Prompt] ──→ [API Call] │ [JSON Response] │ ┌─────────────┼─────────────┐ │ │ Cloze Cards New Word Cards [[word]] → {{c1::word}} word → <b>word</b> │ │ └─────────────┬─────────────┘ │ [Preview Table] │ [User accepts/rejects] │ [Add to Deck]
🤝 Dependencies Anki 2.1.55+ (uses modern Qt6/PyQt6 APIs) google-genai Python package (auto-installed on first run if missing) A free Gemini API key 📝 License
MIT License — feel free to use, modify, and distribute.
