LLM Observability: What Actually Matters

I wrote “LLM Observability: What Actually Matters” to share practical, production-minded guidance on this topic.

What to Track

Latency: P50, P95, P99 response times

Cost: Per query, per user, per day

Quality: Response relevance, accuracy

Errors: Rate limits, failures, timeouts

How to Track It

import structlog

logger = structlog.get_logger()

def track_llm_call(func):
    async def wrapper(*args, **kwargs):
        start = time.time()
        
        try:
            result = await func(*args, **kwargs)
            
            logger.info("llm_call",
                model=result.model,
                tokens=result.usage.total_tokens,
                latency=time.time() - start,
                cost=calculate_cost(result.usage)
            )
            
            return result
        except Exception as e:
            logger.error("llm_call_failed",
                error=str(e),
                latency=time.time() - start
            )
            raise
    
    return wrapper

What to Alert On

• Latency > 5 seconds (P95)
• Error rate > 5%
• Daily cost > budget threshold
• Quality score drop > 10%

Tools That Help

• Application Insights for Azure
• Datadog for multi-cloud
• Custom dashboards with Grafana
• LangSmith for prompt tracking

The Key Insight

You can’t improve what you don’t measure. Start logging everything from day one.\n\n## Takeaways\n\nAdd a concise, personal takeaway and recommended next steps here.\n