A failure mode is the specific model or system behavior behind a business symptom—not simply “accuracy is low.” Naming it helps identify whether a production problem is worth fixing and what to change. section 1.5
Common production failures point to different checks and fixes:
- Retrieved-but-not-used: the right chunk is returned, but the model ignores it. Require answers to use only the context; test with deliberately wrong context, and change models if retrieval is still ignored. section 4.5.6
- Wrong chunk: the retrieved passage lacks the answer. Check whether the answer is in the index; revisit chunking if absent, or hybrid search and reranking if present but ranked too low. section 4.5.6
- Hallucination on retrieval gaps: with no useful context, the model invents an answer. Instruct it to say “I don’t know” and add a downstream check, such as retrieval confidence scoring or a rule-based filter. section 4.5.6
- Latency creep: a growing corpus slows the index and lengthens prompts. Monitor retrieval, prompt assembly, and inference latency; cap retrieved context length. section 4.5.6
Measure failures quantitatively, decide whether they merit attention, and choose the cheapest fix that addresses the behavior. Adaptation is premature if the failure mode cannot be named; it also requires an owner after launch and repeated evaluation when the base model changes. section 1.5 section 4.5.6





