A marketing team at a retail company is using a generative AI model on Vertex AI to produce product descriptions from short bullet lists. They observe that the model's outputs are fluent but frequently invent specifications, such as claiming a jacket is waterproof when no such attribute was provided. The team wants to reduce these fabrications without retraining the model. Which approach should they take?
Grounding the model with the supplied bullet points as explicit context, combined with an instruction to rely solely on that information, constrains generation to the provided facts. This directly reduces fabricated attributes without retraining. It is the standard prompt-level technique for improving factual consistency in Vertex AI generative AI applications when the source data is available at inference time.
Why this answer
Grounding the generation in the provided bullet points by including them as context and instructing the model to use only that information is the most direct way to prevent invented specifications. It leverages the source data already available at inference time and requires no retraining. Parameters such as temperature, top-p, and output length affect style and randomness, not factual fidelity to a given source.
Exam trap
The trap here is assuming that tuning decoding parameters like temperature or top-p will eliminate hallucinations, when those settings only change randomness and cannot supply missing factual grounding.