Vector Voyage: RAG Card Quest – Game Instructions

Game Name: Vector Voyage: RAG Card Quest

Game Description:
Vector Voyage: RAG Card Quest is an immersive, low-tech card game for ages 12+ that takes you on a thrilling journey through the world of Large Language Models (LLMs). Players embark on a quest to chart a ‘vector map’ (knowledge space) by deploying sentence cards to forge paths between words, mimicking AI training. Brace for epic ‘hallucinations’—wild detours like a cow devouring fish!—as you challenge your map with query quests. Then, summon ‘RAG reinforcements’ (external cards) for Retrieval-Augmented Generation, merging maps to conquer errors and unlock accurate discoveries. Perfect for resource-limited settings: just print cards and grids! Designed for 2-6 adventurers or teams, 30-45 minutes, blending cooperative strategy with lessons on AI concepts like biases, prompts, and ethical tech in an adventurous, quest-driven format.

Objective of the Game:
Embark on a quest to understand how LLMs work! Build a “vector map” with words and connections from sentences. Experience “hallucinations” (wrong connections) and use RAG to fix them by combining external knowledge. Win by correcting hallucinations as a group and “unlocking” correct answers.

Game Rules (Step by Step):

  1. Setup (5 min):
    • Each player gets 5 base sentence cards.
    • Draw an empty 5×5 grid in the center (vector map).
    • Set aside the RAG cards and query cards.
  2. Phase 1: Training – Build the Base Vector Space (15-20 min, 25 Epochs/Rounds):
    • Goal: Simulate LLM training by creating a vector space with relations from sentences.
    • Turns: Draw a sentence card (e.g., “cow is herbivore”). Split the words and place them on the grid (freely, but near related words). Draw lines between words in the sentence (neighborhoods).
    • Repeat 25 times (speed up: play in batches of 5 rounds). Skip stopwords like “is” for connections.
    • Educational: Discuss: “See how relations form? This is like embeddings in LLMs – words close if they often appear together.”
  3. Demo Phase: Test the Base (5 min):
    • Draw a query card (e.g., “What does a cow eat?”).
    • Trace a path on the grid (e.g., cow → drinks → water → fish). Discuss the output as a group: If the path is absurd or illogical (e.g., “Cow eats fish?”), it’s a hallucination.
    • Educational: “This is a hallucination – the AI makes up wrong connections due to limited data.”
  4. Phase 2: RAG – Combine with External Knowledge (10 min):
    • Goal: Fix hallucinations without retraining the base.
    • Build a separate 3×3 mini-grid with RAG cards (external sentences, e.g., “herbivore eats grass”).
    • Combination: Draw temporary lines from base to RAG where overlaps exist (e.g., “herbivore” in base links to “eats grass” in RAG). Base stays unchanged!
    • Repeat the query: Trace a new path in the combination (e.g., cow → herbivore → eats → grass). Discuss the improved output as a group: If logical, it’s correct!
    • Educational: “RAG retrieves external info and combines it – hallucinations gone, answers improved!”
  5. End :
    • Discussion Round (5 min): “What did you learn about LLMs? How does RAG make AI smarter?”