AWS AI Practitioner Question 33

Key Takeaways

AWS AI practitioner exam prep question 33. A marketing team wants to generate product descriptions in multiple languages using Amazon Bedrock. They've noticed that the model sometimes produces text that's too long, includes competitors names, and occasionally hallucinates product features that don't exist. [music] Which combination of techniques should the team to use to address all three issues? So, let's identify the key hint. First, text too long, so we need output length control. Two, includes competitor names, so we need content filtering. Three, it hallucinates features, so we need a grounding in real product data. So, option one, fine-tune the foundation model on company product cataloges. Option two, use bedrock guard rails for word filtering, set max tokens, and inference parameters, and implement rag with a product database. Option three, switch to a smaller model with shorter default outputs. And then option [music] four, add system prompts instructing the model to avoid competitors and limit length. Add your answer in the comments below. So, the hint gives us three separate problems and we actually kind of need three separate solutions. So, option one, fine-tuning is expensive and doesn't directly solve link control or competitor name filtering. Option three, smaller models don't solve hallucination or competitor filtering. They might also reduce output quality. Option four, system prompts help but aren't reliable for strict enforcement. The model can still ignore link limits and mention competitors. So, the correct answer is actually option two. Guard rails plus max tokens plus rag. It's actually three different solutions. So max tokens controls the output length at the API level. Guard rail word filters are going to block competitors names reliably. And then retrieval augmented generation or rag is going to ground the model in actual product data that the model can read. So there's no hallucinated features. Each problem basically gets a purpose-built solution. Three problems, three tools. Inference parameters for length, guardrails for filtering, rag for grounding. Are you ready to pass the exam? Visit aws.co.com. co-lab.com.