Quest 22 - Hallucination Prevention System

Quest 22: Hallucination Prevention System

hard 35 minutes

---

🎯 Learning Objectives

• ✅ Understand hallucination as a fundamental LLM challenge requiring layered defense
• ✅ Implement multiple prevention strategies: temperature, context injection, citation, confidence
• ✅ Build a robust system that layers strategies instead of relying on one
• ✅ Detect sparse context and warn users about potential hallucination risk

---

📖 Concept: Hallucination Prevention

LLM มีนิสัย “แต่งเรื่อง” — เรียกว่า hallucination คือสร้างข้อมูลที่ดูสมเหตุสมผลแต่ไม่จริง การป้องกันไม่ใช่แค่ technique เดียว แต่ต้องใช้ defense in depth — หลายชั้นซ้อนกัน

แต่ละชั้นมีจุดแข็งและจุดอ่อนต่างกัน:

• Temperature reduction: ลดความ creative ของ model
• Context injection: ให้ข้อมูลจริงเป็น foundation
• Citation requirement: บังคับให้ model อ้างอิงแหล่งที่มา
• Confidence check: เตือนเมื่อ context ไม่พอ

เหมือน martial arts defense: ไม่ได้มีแค่ท่า block เดียว — คุณใช้ footwork + guard + head movement ซ้อนกัน

---

⚙️ How It Works

The Multi-Layer Prevention System

1. Receive prompt, context, and options
         ↓
2. Apply Layer 1: Temperature check
   └── Warn if temperature > 0.7 (too creative)
         ↓
3. Apply Layer 2: Context injection
   └── Enhance prompt with retrieved context
         ↓
4. Apply Layer 3: Citation requirement
   └── Add citation instructions when enabled
         ↓
5. Apply Layer 4: Confidence check
   └── Warn if context is sparse or empty
         ↓
6. Return { safe, warnings, enhancedPrompt }

Warning Levels

Warning	Trigger	Action
Temperature	> 0.7	Suggest lowering to 0.3-0.5
Context sparse	< threshold chunks	Warn about hallucination risk
No context	Empty context array	Strong warning, low confidence
Citation needed	requireCitation = true	Add citation instructions

---

💡 Example: Building the System

Step 1: Basic structure

function preventHallucination(prompt, context, options) {
  const { temperature = 0.5, requireCitation = false, confidenceThreshold = 0.5 } = options;
  const warnings = [];

  // Layer 1: Temperature check
  if (temperature > 0.7) {
    warnings.push('High temperature may increase hallucination risk. Consider lowering to 0.3-0.5.');
  }

  // Layer 2: Context injection
  let enhancedPrompt = prompt;
  if (context.length > 0) {
    enhancedPrompt = `Based on the following context:\n${context.join('\n---\n')}\n\nQuestion: ${prompt}`;
  }

  // Layer 3: Citation requirement
  if (requireCitation) {
    enhancedPrompt += '\n\nPlease cite your sources for each claim.';
  }

  // Layer 4: Confidence check
  if (context.length === 0) {
    warnings.push('No context provided. Response may contain hallucinated information.');
  } else if (context.length < confidenceThreshold * 5) {
    warnings.push('Limited context available. Confidence in response may be low.');
  }

  const safe = warnings.length === 0;
  return { safe, warnings, enhancedPrompt };
}

Step 2: Verify all layers work

node test.js
# Test: returns required structure
# Test: warns on high temperature
# Test: adds citation instructions
# Test: warns on empty context
# Test: considers context quantity
# Test: applies multiple strategies

---

⚠️ Common Mistakes

Mistake 1: Only implementing one strategy

“Temperature reduction is enough” → No single strategy catches everything. Layer multiple approaches.

Mistake 2: Not returning warnings

Returning { safe, enhancedPrompt } without warnings array → Warnings help users understand WHY the system flagged something.

Mistake 3: Ignoring context quantity

Treating 1 context chunk the same as 100 → Fewer chunks = lower confidence = more warnings.

Mistake 4: Making enhancedPrompt too long

Injecting massive context into every prompt → Keep context focused and relevant. Quality > quantity.

---

📝 Knowledge Check

📝 Knowledge Check

Q1:What is 'defense in depth' in the context of hallucination prevention?

Using only the strongest prevention techniqueLayering multiple strategies so no single point of failure causes hallucinationPreventing hallucination at the model training level onlyUsing a single temperature setting for all prompts

Q2:What warning should the system generate when temperature is high?

'Model is running too fast''High temperature may increase hallucination risk. Consider lowering to 0.3-0.5''Temperature is optimal''No warnings needed'

Q3:What happens when context is empty in a hallucination prevention system?

The system proceeds normallyThe system returns a strong warning about potential hallucinationThe system crashesThe system generates context automatically

---

🏋️ Quest: Hallucination Prevention System

Now it’s time to build a multi-layer defense system!

1. Download ไฟล์เริ่มต้นของ quest:

   npx bluebeltdojo download quest-22-hallucination-prevention
   cd quest-22-hallucination-prevention

2. เปิด problem.js ใน editor ของคุณพร้อมความช่วยเหลือของ AI

3. Implement preventHallucination(prompt, context, options) — ใช้ 4 prevention layers

4. สำคัญ: ทดสอบทุก layer — temperature, context, citation, confidence

5. ตรวจสอบ solution ของคุณ:

   node test.js

6. When all tests pass, submit your solution:

   npx bluebeltdojo submit

💡 Tip: Hallucination prevention คือ arms race — ยิ่ง model เก่งขึ้น ยิ่งต้องมี defense ซับซ้อนขึ้น

---

คำใบ้

• อ่าน instructions ใน problem.js อย่างละเอียด
• ทดสอบทุก layer: temperature warning, context injection, citation instruction, confidence check
• ถ้าติดขัด ลองอ่าน “Common Mistakes” อีกครั้ง — อย่าดู solution โดยตรง