Diagnosing System Failures
Explore how to systematically diagnose and fix wrong outputs in AI systems by distinguishing prompt failures, hallucinations, and model mismatches. Understand the importance of both semantic and structural validation and learn a step-by-step framework for root-causing issues and avoiding common diagnostic mistakes.
The loop from the previous lesson, evaluate, diagnose, fix, re-verify, only works if the diagnosis step actually identifies the real cause. A wrong output can come from several different places, and each one demands a different fix. Rewriting a prompt does nothing for a model that has genuinely exceeded its capability on this task. Adding more retrieved context does nothing for an instruction that was simply ambiguous. Getting the diagnosis wrong does not just fail to fix the problem, it burns an iteration on a change that was never going to help.
In this lesson, we will cover:
Distinguishing three common causes of a wrong output, prompt failure, hallucination, and model mismatch, and the different fix each one actually needs
Semantic validation versus structural validation, and why passing one says nothing about the other
Root-causing a regression systematically, using a document refresh that breaks retrieval as a worked example
A diagnostic framework for narrowing from a symptom to an actual, fixable cause ...