Swedish-Style Crossword Puzzle Generator

A high-performance Java-based puzzle generator with theme-based word filtering and daily automated generation.

Features

• Swedish-style crossword puzzles with arrow clues
• Theme-based word filtering using semantic similarity graph
• Daily automated generation via Docker + cron
• JSON export format compatible with web frontends
• Genetic algorithm for optimal grid layouts
• Constraint satisfaction for word placement

Architecture

Components

1. SwedishGenerator.java - Core puzzle generation engine

   • Genetic algorithm for mask generation
   • CSP solver for word filling
   • Optimized for Dutch word lists

2. ThemeGraph.java - Theme-based word scoring system

   • Predefined theme keywords (news, tech, sports, etc.)
   • Edit distance similarity matching
   • Automatic theme detection

3. DailyGenerator.java - Daily puzzle automation

   • Generates themed puzzles
   • JSON output with metadata
   • Index file generation

4. ExportFormat.java - Export to standard format

   • Grid cropping and optimization
   • Arrow cell calculation
   • Compatible with existing frontends

Usage

Local Development

# Compile
./compile.sh

# Run Main (interactive)
java -cp ~/dev/.target puzzle.Main --seed 42 --pop 18 --gens 100

# Generate daily puzzles
java -cp ~/dev/.target puzzle.DailyGenerator

Docker Deployment

# Build image
docker build -t puzzle-generator .

# Run with docker-compose
docker-compose up -d puzzle_gen_java

# View logs
docker logs -f puzzle_gen_java

Environment Variables

Variable	Default	Description
OUT_DIR	/data/puzzles	Output directory for generated puzzles
PUZZLES_PER_DAY	3	Number of puzzles to generate daily
WORDS_PATH	./word-list.txt	Path to word list file
THEME_FILTER	true	Enable theme-based word filtering
THEME_MIN_SCORE	0.6	Minimum theme score (0.0-1.0)
LM_STUDIO_BASE_URL	-	LM Studio URL (future feature)
GENERATE_ON_START	false	Generate puzzles on container startup

Theme System

Supported Themes

• algemeen - General/common words
• nieuws - News/politics
• technologie - Technology
• sport - Sports
• weer - Weather/nature
• economie - Economy
• gezondheid - Health

Theme Filtering

Words are scored against themes using:

1. Direct matching - Word is in theme keyword list (score: 1.0)
2. Substring matching - Partial word overlap (score: 0.7)
3. Edit distance - Fuzzy matching for variations (score: 0.8-0.9)

Example:

ThemeGraph.filterByTheme(words, "technologie", 0.6);
// Returns: COMPUTER, INTERNET, SOFTWARE, DATA, etc.

Output Format

Puzzle JSON

{
  "date": "2025-12-19",
  "theme": "technologie",
  "difficulty": 1,
  "rewards": {
    "coins": 50,
    "stars": 2,
    "hints": 1
  },
  "gridv2": [
    "###COMPUTER###",
    "#I#O#E#E#O#"
  ],
  "words": [
    {
      "word": "COMPUTER",
      "clue": "COMPUTER",
      "startRow": 0,
      "startCol": 3,
      "direction": "horizontal",
      "answer": "COMPUTER",
      "arrowRow": 0,
      "arrowCol": 2
    }
  ]
}

Index JSON

{
  "date": "2025-12-19",
  "files": [
    "crossword_2025-12-19_01_technologie.json",
    "crossword_2025-12-19_02_sport.json",
    "crossword_2025-12-19_03_nieuws.json"
  ]
}

Scheduling

Puzzles are generated daily at 3:15 AM (configurable in crontab).

Edit crontab to change schedule:

# Daily at 3:15 AM
15 3 * * * java -cp /app/target puzzle.DailyGenerator

# Every 6 hours
0 */6 * * * java -cp /app/target puzzle.DailyGenerator

# Weekly on Monday at 1 AM
0 1 * * 1 java -cp /app/target puzzle.DailyGenerator

Word List Format

Plain text file, one word per line, uppercase A-Z only, 2-8 characters:

EU
UUR
AUTO
BOOM
COMPUTER
INTERNET
...

Performance

• Mask generation: ~2-5 seconds (genetic algorithm)
• Word filling: ~5-30 seconds (CSP solver with MRV heuristic)
• Total per puzzle: ~10-40 seconds

Optimizations:

• Positional indexing for fast candidate lookup
• Sorted intersection for constraint checking
• No large array allocations during search
• Progress bar with real-time stats

Integration with LM Studio (Future)

The system is prepared for LM Studio integration to generate themed clues:

docker-compose up -d
# Set LM_STUDIO_BASE_URL in docker-compose.yml
# Container will query LM Studio for contextual clues based on themes

This will enhance clues from simple word repetition to semantic hints.

Migration from Node.js

The Java version maintains module-wise compatibility with the Node.js generator:

Node.js	Java
swedish_generator.js	SwedishGenerator.java
export_format.js	ExportFormat.java
main.js	Main.java + DailyGenerator.java
N/A	ThemeGraph.java (new)

Volume Management

Puzzles are stored in a Docker volume outside the workspace:

# Default location
/var/lib/puzzle-data

# Custom location
export PUZZLE_OUTPUT_DIR=/path/to/puzzles
docker-compose up -d

# View puzzles
ls -lh /var/lib/puzzle-data/*.json

Troubleshooting

No puzzles generated

• Check word list has enough words (minimum 50)
• Lower THEME_MIN_SCORE if using theme filtering
• Increase PUZZLES_PER_DAY attempts

Container not starting

docker logs puzzle_gen_java
# Check for compilation errors or missing files

Low quality puzzles

• Increase --gens parameter (more genetic iterations)
• Increase --pop parameter (larger population)
• Ensure word list has good variety of lengths 2-8

License

MIT

Authors

Original Node.js version + Java port with theme system