Basic concepts of generative AI (GenAI)

Generative Image Synthesis (Option A) fulfills the requirement to automatically create photorealistic thumbnails from text titles (e.g., using Amazon Titan Image Generator). Generative Video Synthesis (Option C) fulfills the requirement to automate the creation of 30-second news highlight reels. Conversational AI Assistants (Option E) fulfill the requirement for a virtual agent that site visitors can interact with to find articles. Source Code Generation (Option B) writes software, not thumbnails. Automated Audio Transcription (Option D) converts speech to text, which is not required for images, video reels, or chat. Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/titan-image-models.html

Real-Time Data Processing (Option D) is critical for agents operating in highly dynamic environments (like network routing). While a Large Language Model's parametric knowledge is static based on its last training run, coupling the agent with real-time data processing (via Action Groups or API tool use) allows the system to base its reasoning and orchestration on the exact, current state of the environment. Fine-tuning (Option B) is a slow, batch process that cannot be done in real-time. Memory Management (Option A) stores past conversation context, not live telemetry. Semantic search (Option C) retrieves static documents. Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html

The conversation history has exceeded the model's context window (Option B). The context window is the maximum amount of text (tokens) an LLM can remember during a single request. In a chat application, if the history of the conversation grows larger than this window, the oldest messages (like initial instructions) are dropped, causing the model to 'forget.' Lowering temperature (Option A) reduces randomness, but cannot fix forgotten text. Grounding data (Option C) improves facts, it doesn't cause amnesia. Data drift (Option D) affects traditional ML accuracy over time, not short-term chat memory. Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters.html

Implementing a preprocessing script (Option A) reduces the size of the transcript, thereby lowering input tokens, reducing cost, and speeding up the model's reading time. Restricting the maximum output tokens (Option D) forces the model to stop generating earlier, reducing output token costs and cutting down the sequential generation latency. SageMaker Savings Plans (Option B) do not apply to serverless Amazon Bedrock API calls. Top-K (Option C) determines the sampling pool for next-word prediction (randomness), not processing speed or cost. Model weights (Option E) are managed by AWS on Bedrock and cannot be cached by the customer in ElastiCache. Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters.html

Prompt Engineering (Option D) is the practice of carefully designing, refining, and optimizing the input text (prompts) given to a Generative AI model to guide it toward producing more accurate, relevant, and properly formatted outputs, entirely without changing the model's weights. Parameter-Efficient Fine-Tuning (Option A) involves actually updating a small subset of the model's weights, which the team explicitly did not do. Text Tokenization (Option B) is how the model parses words. Semantic Chunking (Option C) is the process of breaking large documents into smaller pieces for a vector database. Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-engineering-guidelines.html

Generating output tokens is sequential and adds latency (Option B). In autoregressive language models, output tokens are generated one by one. The more tokens the model has to generate (e.g., an 800-word essay), the longer the user has to wait for the complete response. It is not an issue of exceeding the context window (Option A), as that would result in an error, not just latency. Foundation models are static during inference; their weights are not updated in real-time (Option C). API Gateway throttling results in 429 HTTP errors, not slow, successful generation (Option D). Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters.html

Enforcing a strict limit on the max_tokens parameter (Option A) prevents users from generating excessively long 5,000-word articles, directly capping output token costs. Pre-processing the user's prompt (Option E) condenses verbose background notes, reducing the number of input tokens sent to the more expensive, main foundation model. AWS Fargate (Option B) is a container compute engine, not a model hosting platform. Learning rates (Option C) are used during model training, not during inference. Deploying on a larger EC2 instance (Option D) involves self-hosting, which increases fixed infrastructure costs and does not solve the root problem of massive token generation. Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters.html

A Multi-Agent System (Option C) is characterized by a network of distributed, autonomous agents (the building AI entities) that communicate, collaborate, or negotiate to solve a larger, decentralized problem (city-wide energy optimization). Transfer Learning (Option A) involves porting weights from a pre-trained model to a new task. Unsupervised Clustering (Option B) groups unlabeled data. While the agents might run on local hardware (Edge Inferencing, Option D), the collective negotiation and collaboration architecture itself is definitively a Multi-Agent System. Reference: https://aws.amazon.com/what-is/multi-agent-systems/

Data Selection (Option A) ensures the proprietary data is curated and scrubbed of PHI before any processing occurs. Fine-tuning (Option B) adapts the licensed open-source base model to the highly specialized medical records domain. Evaluation (Option E) ensures the adapted model is rigorously tested against diagnostic criteria before use. Pre-training (Option D) is incorrect because the provider licensed an existing open-source model rather than building one from scratch. Distributed Training (Option C) is an infrastructure scaling technique, not a conceptual phase of the ML lifecycle. Reference: https://docs.aws.amazon.com/sagemaker/latest/dg/jumpstart-foundation-models.html