Chapter 3: LLM-Specific Challenges

1. Hallucination Management

LLMs can generate plausible but incorrect information:

class HallucinationDetector:
    def __init__(self, fact_checker_model):
        self.fact_checker = fact_checker_model
    
    def check_response(self, query: str, response: str, sources: list) -> dict:
        """Check if response is grounded in sources."""
        
        # Extract claims from response
        claims = self.extract_claims(response)
        
        # Verify each claim against sources
        verified_claims = []
        hallucinated_claims = []
        
        for claim in claims:
            if self.verify_against_sources(claim, sources):
                verified_claims.append(claim)
            else:
                hallucinated_claims.append(claim)
        
        return {
            "hallucination_rate": len(hallucinated_claims) / len(claims),
            "verified_claims": verified_claims,
            "hallucinated_claims": hallucinated_claims
        }

2. Safety and Content Filtering

class SafetyGuard:
    def __init__(self):
        self.blocked_topics = ["illegal activities", "harmful content", ...]
        self.pii_patterns = [r"\b\d{3}-\d{2}-\d{4}\b", ...]  # SSN, etc.
    
    def check_input(self, user_input: str) -> tuple[bool, str]:
        """Check user input for safety issues."""
        # Topic filtering
        for topic in self.blocked_topics:
            if self.detects_topic(user_input, topic):
                return False, f"Blocked topic: {topic}"
        
        return True, "OK"
    
    def check_output(self, llm_output: str) -> str:
        """Sanitize LLM output."""
        # Remove PII
        for pattern in self.pii_patterns:
            llm_output = re.sub(pattern, "[REDACTED]", llm_output)
        
        return llm_output

3. Latency Optimization

Technique	Latency Reduction	Trade-off
Streaming	Perceived latency	Implementation complexity
Caching	90%+ on cache hits	Storage, cache invalidation
Smaller models	50-80%	Quality
Batching	Variable	Response time

4. Cost Management

class CostTracker:
    def __init__(self, budget_limit: float):
        self.budget_limit = budget_limit
        self.total_cost = 0
        
    def track_request(self, model: str, input_tokens: int, output_tokens: int):
        cost = calculate_cost(input_tokens, output_tokens, model)
        self.total_cost += cost
        
        if self.total_cost > self.budget_limit * 0.8:
            alert_team("Approaching budget limit")
        
        if self.total_cost > self.budget_limit:
            raise BudgetExceededError("Budget limit reached")
        
        return cost

---

Next: Module 10 Quiz