Convex AI Card Generation

Convex AI Card Generation

Overview

This skill handles AI-powered card generation for PictionAI's Pictionary gameplay. Cards are generated dynamically using Groq LLM with Vercel AI SDK's generateText function, including structured output validation via Zod schemas and automatic fallback to hardcoded card libraries when API limits are reached.

Architecture

Generation Flow

1. Request card for category
   ↓
2. Check card availability/balance
   ↓
3. Try Groq (primary) → Fallback to other models → Fallback to hardcoded library
   ↓
4. Validate with Zod schema
   ↓
5. Return structured card object { word, description, difficulty, category }

AI Models (Priority Order)

1. groq-mixtral-8x7b-instruct (Primary) - Best quality, lower latency
2. groq-llama2-70b-chat (Secondary) - If Mixtral unavailable
3. groq-llama-3.1-70b (Tertiary) - Fallback option
4. Hardcoded Library (Final) - When all API models fail

Key Action

generateCards

Generate one or multiple drawing cards for a specific category with AI.

action generateCards {
  args: {
    category: "animals" | "objects" | "actions" | "movies" | "sports",
    count?: number  // Default: 1, Max: 5
  }
  // Returns:
  // {
  //   cards: Array<{
  //     word: string,
  //     description: string,
  //     difficulty: "easy" | "medium" | "hard",
  //     category: string
  //   }>,
  //   source: "ai" | "fallback"  // Indicates if AI or hardcoded
  // }
}

Card Data Structure

interface Card {
  word: string; // Single word to draw (required)
  description: string; // Drawing hints/clues (1-100 chars)
  difficulty: "easy" | "medium" | "hard";
  category: string; // Category name
  created_at?: number; // Timestamp
  ai_generated?: boolean; // True if from Groq
}

Supported Categories

• animals - Creatures, pets, wildlife
• objects - Inanimate items, tools, furniture
• actions - Verbs, activities, gestures
• movies - Film titles, characters, scenes
• sports - Sports, games, athletic activities
• food - Dishes, ingredients, cuisine
• technology - Gadgets, software, digital items

Implementation Details

Groq Configuration

const groq = new Groq({
  apiKey: env.GROQ_API_KEY, // Environment variable
  baseURL: "https://api.groq.com/openai/v1",
});

Zod Schema for Validation

const cardSchema = z.object({
  word: z.string().min(1).max(30),
  description: z.string().min(10).max(100),
  difficulty: z.enum(["easy", "medium", "hard"]),
  category: z.string(),
});

const cardsResponseSchema = z.object({
  cards: z.array(cardSchema),
});

AI Prompt Template

Generate {count} unique Pictionary drawing cards for the "{category}" category.
For each card, provide:
- word: A single word to draw
- description: A brief hint (1-100 characters)
- difficulty: easy, medium, or hard
- category: {category}

Format as JSON array. Make words creative and diverse.

Fallback Card Libraries

When all AI models fail, the system falls back to hardcoded libraries:

const cardLibraries = {
  animals: [
    {
      word: "elephant",
      description: "Large animal with long trunk",
      difficulty: "easy",
      category: "animals",
    },
    {
      word: "giraffe",
      description: "Tall with spotted pattern",
      difficulty: "easy",
      category: "animals",
    },
    // ... more animals
  ],
  objects: [
    {
      word: "bicycle",
      description: "Two-wheeled vehicle",
      difficulty: "easy",
      category: "objects",
    },
    // ... more objects
  ],
  // ... other categories
};

React Integration

// In game setup, generate initial card
const generateCardsAction = useAction(api.actions.generateCards);

async function setupGameCards(category: string) {
  const result = await generateCardsAction({
    category,
    count: 5, // Generate 5 cards per round
  });

  console.log(`Cards generated from ${result.source}`);
  return result.cards;
}

Error Handling

Graceful Degradation

1. API Quota Exceeded → Automatically switch to fallback
2. Network Error → Retry once, then use fallback
3. Invalid Response → Log error, use fallback
4. Schema Validation Fail → Discard AI response, use fallback

Logging

console.log(`Generated cards from ${source} for category: ${category}`);
// source: "ai" or "fallback"

Performance Considerations

• AI Generation: ~500-1000ms per card (includes API latency)
• Fallback Lookup: ~50ms per card
• Caching: Consider caching generated cards per category/session
• Batch Generation: Generate 5-10 cards at game start, not on-demand per turn

Best Practices

Do's

✅ Generate multiple cards at game start (not per turn) ✅ Cache cards for session duration ✅ Log both AI and fallback usage for monitoring ✅ Validate all AI responses against Zod schema ✅ Handle quota errors gracefully

Don'ts

❌ Don't wait for card generation on each turn ❌ Don't expose Groq API key in frontend ❌ Don't retry forever on network errors ❌ Don't mix AI and fallback for same category in single game

Example: Complete Card Setup

// Game initialization
export const initializeGameCards = action({
  args: { game_id: v.id("games"), category: v.string() },
  handler: async (ctx, args) => {
    const generateCards = await ctx.runAction(api.actions.generateCards, {
      category: args.category,
      count: 10,
    });

    // Store cards in game doc
    await ctx.runMutation(api.mutations.games.storeGameCards, {
      game_id: args.game_id,
      cards: generateCards.cards,
    });

    return {
      count: generateCards.cards.length,
      source: generateCards.source,
    };
  },
});

See Also

• convex/actions/generateCards.ts - Complete action implementation
• Groq API docs: https://console.groq.com/docs
• Vercel AI SDK: https://sdk.vercel.ai