Design considerations for applications that u..

Fine-tuning the foundation model (Option B) is required to teach the model the deep semantic nuances, tone, and proprietary precedents unique to the firm, which standard models and basic prompting fail to capture. Implementing Retrieval-Augmented Generation (RAG) (Option D) is required to satisfy the need to cross-reference against live, daily federal court rulings without constantly retraining the model. Pre-training from scratch (Option A) is excessively expensive. Model distillation (Option C) does not add legal knowledge. Zero-shot prompting (Option E) will fail to capture the deep proprietary knowledge required. Reference: https://aws.amazon.com/blogs/machine-learning/customize-models-in-amazon-bedrock-with-your-own-data-using-fine-tuning-and-continued-pre-training/

Model complexity and reasoning capability (Option B) (often correlated with larger parameter sizes) is the most relevant criterion. Processing highly intricate data matrices and extracting deep logical patterns requires a highly capable, complex model. Smaller, simpler models would fail to grasp the nuanced correlations in the user data. Audio/image modalities (Option A) are for media generation, not text data reasoning. Prompt caching (Option C) optimizes cost, and edge deployment (Option D) is for IoT devices, neither of which addresses the need for advanced reasoning.

Increasing the Temperature to 0.9 (Option B) flattens the probability distribution, making the model more likely to select unexpected and creative tokens rather than just the most common ones. Increasing Top-P to 0.9 (Option D) widens the 'nucleus' of available tokens, allowing the model to choose from a massive pool of vocabulary, further enhancing diversity and creativity. Decreasing Temperature (Option A), Top-P (Option C), or Top-K (Option E) all severely restrict the model's choices to only the most probable, safe, and repetitive words, which is the exact opposite of the creative directors' goal. Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters.html

RAG drastically reduces hallucinations (Option A) by providing the model with exact, factual grounding documents to read from, rather than relying on its internal memory. RAG also eliminates the need for expensive, frequent retraining (Option D) because new knowledge is simply added to the search database, not baked into the model's weights. RAG does not increase hard context window limits (Option B). It does not convert unstructured text into SQL (Option C). Generative AI remains probabilistic, so RAG does not make outputs mathematically verifiable or strictly deterministic (Option E). Reference: https://aws.amazon.com/what-is/retrieval-augmented-generation/

To overcome a model's 'knowledge cutoff' date without retraining it, the platform must implement a Retrieval-Augmented Generation (RAG) pipeline (Option A). Utilizing Amazon Bedrock Knowledge Bases (Option C) provides a fully managed way to implement this RAG pipeline by indexing the new product catalog. Increasing Temperature (Option B) simply increases hallucinations and guessing. Pre-training (Option D) violates the requirement to not train a new model from scratch and is extremely expensive. Decreasing Top-P (Option E) restricts vocabulary but doesn't provide new facts. Reference: https://aws.amazon.com/what-is/retrieval-augmented-generation/

Inference latency (Option A) is the time it takes for a model to process the input and return a response. For a real-time mobile application analyzing live video frames, low latency is critical to prevent UI lag and maintain a seamless user experience. Fine-tuning (Option B), context windows (Option C), and parameter counts (Option D) are secondary to the sheer speed required for real-time video processing applications.

Lowering the Temperature to near 0 (Option A) removes randomness, forcing deterministic, factual extraction. Decreasing the Top-P value (Option B) further narrows the token selection pool, ensuring the model only considers the absolute most likely medical codes. Configuring a strict Maximum Output Tokens limit (Option D) caps the response length, minimizing costs and cutting off potential conversational filler text. Increasing Top-K (Option C) or Temperature (Option E) would introduce unwanted creativity, variability, and potential hallucinations into a strict medical coding pipeline. Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters.html

Model customization and fine-tuning options (Option A) are critical because the firm needs to alter the model's weights using their highly proprietary trading algorithms (a capability not all managed models support). Inference latency (Option B) is vital to ensure minimal delay during active, fast-paced trading hours. Model reasoning complexity (Option C) ensures the model has the capacity for deep mathematical and logical analysis. Image generation (Option D) and translation (Option E) are irrelevant to numerical financial analytics.

Agentic AI architectures empower models to orchestrate complex workflows by connecting to external systems via APIs (Option A). A standard RAG architecture is strictly 'read-only'—it can retrieve documents and generate text, but it cannot execute a trade or update a database. Agentic AI is not immune to hallucinations (Option B is false). Standard RAG can process numbers (Option C is false). Agents still rely heavily on prompt engineering (specifically system prompts outlining their persona and available tools) to function (Option D is false). Reference: https://aws.amazon.com/what-is/ai-agents/