📖 Practical Task: AI Q&A System

Welcome! 👋 In this exercise, you will implement a simple AI-powered Q&A system using the provided dataset. This is designed to test your ability to design, code, and think about AI systems.

---

🔹 Files Provided

• qa_mixed_expanded.json → Dataset of Q&A pairs (includes:
  • Regular questions (answerable)
  • Trick questions (should escalate)
  • Borderline paraphrases (test similarity threshold))
• main.py → Starter code (JSON loading, embeddings model, and placeholders for you).

---

🔹 Your Tasks

1. Implement answer(query)

• Input: a text query.
• Logic:
  • Encode the query.
  • Compute similarity against dataset questions.
  • If similarity score > threshold → return the corresponding answer.
  • Else → return "escalate".
• Default threshold: 0.6 (you may adjust if needed).

2. Test Your System

Run main.py and try queries:

• ✅ Regular: “What is the capital of France?” → should return "Paris".
• ⚠️ Trick: “Who is the CEO of OpenAI?” → should return "escalate".
• 🤔 Borderline: “What’s the capital city of France?” → should still return "Paris".

3. (Bonus) Implement Feedback Loop

Add a function:

def add_feedback(query: str, correct_answer: str):
    # Append a new Q&A pair to dataset

• Purpose: capture corrected answers from users.
• Ensure safe handling (don’t overwrite data blindly).

4. Explain Your Approach

After coding, be prepared to explain:

• Why you chose your similarity threshold.
• How you would scale this system for larger datasets.
• How you would use the feedback loop to “teach” the system safely.

---

🔹 How to Run

python main.py

You can then type queries interactively. Type quit to exit.

---

✅ What We’re Looking For

• Correct handling of regular, trick, and borderline queries.
• Clean, readable code.
• Evidence of system thinking (not just coding).
• Bonus: ability to extend with safe feedback/learning loop.