When the Correct Answer Is Nothing, What Does Your Pipeline Return?

LLM pipelines can fail by confidently inventing answers when no correct information exists in the source material. This issue arises when reliability mechanisms, such as required schema fields for structured outputs or similarity thresholds in semantic caches, remove the system's ability to abstain. For instance, a cache might incorrectly match questions with opposite meanings, while an extraction tool might hallucinate a date to satisfy a schema. While solutions like making fields nullable or requiring evidence for each extracted value are proposed, their implementation can contain subtle flaws that are difficult to catch during review.