Preface
In early 2024, I shipped my first LLM-powered feature to production. It was a summarization pipeline for customer support tickets — straightforward, well-scoped, and impressively effective during development. Within 72 hours of deployment, it had hallucinated a refund policy that did not exist, misattributed a complaint to the wrong product category, and confidently summarized a Spanish-language ticket in English while inventing details that appeared nowhere in the original text.
I had extensive monitoring. Latency dashboards, token-usage graphs, error-rate alerts — the full observability stack. Everything looked green. The system was fast, available, and cheap to run. It was also producing wrong outputs that nobody caught until a support ...
Become an O’Reilly member and get unlimited access to this title plus top books and audiobooks from O’Reilly and nearly 200 top publishers, thousands of courses curated by job role, 150+ live events each month,
and much more.
Read now
Unlock full access