Task 3.1: Describe design considerations for applications that use foundation models (FMs).

• ARaise the temperature so the model explores more word choices

  Incorrect

  A higher temperature flattens the probability distribution and makes lower-probability tokens more likely, which leads to more random output.

• BRaise Top P so that a larger share of candidate tokens is considered

  Incorrect

  A higher Top P widens the pool of candidate tokens and lets the model consider less likely outputs, the opposite of what the reviewers want.

• CLower the temperature so the model favors higher-probability tokens

  Correct

  A lower temperature steepens the probability distribution, makes the model select higher-probability tokens, and leads to more deterministic responses.

• DIncrease the maximum response length so the model has more room

  Incorrect

  Response length only limits how many tokens are returned. It does not change how the model chooses its words.

Temperature, Top K and Top P control randomness; lowering them makes output more focused and repeatable. Length parameters only cap or penalize length.

• AFine-tune the foundation model on the updated policy documents at the start of every month

  Incorrect

  Retraining foundation models for organization-specific information has high computational and financial costs, and the team wants to avoid monthly training.

• BRaise the temperature so the model generalizes to the new policies

  Incorrect

  Temperature only changes how random the output is. It cannot give the model information it has never seen.

• CUse a larger model whose training data is more recent

  Incorrect

  LLM training data is static and has a cut-off date, so a newer model would still miss the company's internal policies and next month's changes.

• DUse Retrieval Augmented Generation with Amazon Bedrock Knowledge Bases

  Correct

  RAG has the model consult an authoritative knowledge base before it answers, without retraining. Knowledge Bases can answer with direct quotations from the sources or with responses generated from them.

Changing facts plus a need for sources is the classic case for RAG: update the documents, not the model.

• AAmazon OpenSearch Service

  Correct

  OpenSearch Service offers k-nearest neighbor (k-NN) search over vectors and is a vector database option for generative AI applications.

• BAWS Glue DataBrew

  Incorrect

  DataBrew is a visual data preparation tool for cleaning and normalizing data. It does not store or search embeddings.

• CAmazon Polly

  Incorrect

  Amazon Polly converts text into lifelike speech. It has nothing to do with storing vectors.

• DAmazon RDS for PostgreSQL with the pgvector extension

  Correct

  Amazon RDS for PostgreSQL and Aurora PostgreSQL support the pgvector extension to store embeddings and perform similarity searches.

• EAmazon Translate

  Incorrect

  Amazon Translate translates text between languages. It does not index embeddings.

On AWS, vectors can live in a dedicated search engine (OpenSearch Service) or next to existing relational data (Aurora or RDS for PostgreSQL with pgvector), among other options.

• AModel distillation from the large model into a smaller model

  Correct

  Distillation transfers knowledge from a larger model (teacher) to a smaller, faster, cost-efficient model (student), improving the student's performance for a specific use case.

• BContinued pre-training of the large model on a large set of unlabeled email text

  Incorrect

  Continued pre-training adapts a model to a domain with unlabeled data. The large model would stay just as slow and costly to run.

• CProvisioned Throughput for the large model

  Incorrect

  Provisioned Throughput buys a higher level of throughput for a model at a fixed hourly cost. It does not make the model smaller or cheaper per request.

• DA higher temperature for the large model's requests

  Incorrect

  Temperature changes the randomness of output, not the model's size, speed or cost.

When a big model's quality is right but its cost and latency are not, distillation captures that quality in a smaller model for the specific task.

• ATranslating one product description from English into Spanish for the online catalog

  Incorrect

  Translation is a single-step task a model completes from one prompt; it needs no planning or actions in other systems.

• BResolving a customer query by asking questions, checking records and deciding on a handoff

  Correct

  AWS describes this contact center case as an AI agent: it asks questions, looks up information, responds with a solution and decides whether to pass the query to a human.

• CClassifying the sentiment of a single customer review

  Incorrect

  Sentiment classification is a single prompt-and-answer task that a model handles without tools or planning.

• DSummarizing one long internal policy document that an employee pastes directly into the prompt

  Incorrect

  Summarizing text supplied in the prompt is a one-step generation task; there is nothing to look up or act on.

Agents earn their extra complexity on multi-step goals that need decisions, tool calls and data lookups, not on single-step generation tasks.

• AFine-tune the model on each uploaded manual before the user's chat session starts

  Incorrect

  Model customization is a training job that adjusts the model's parameters; running one per uploaded manual would add cost and delay, not remove it.

• BSend the questions through Amazon Bedrock batch inference

  Incorrect

  Batch inference processes prompts asynchronously from files in S3. It does not suit an interactive chat session.

• CChoose a model that supports prompt caching and cache the manual

  Correct

  Prompt caching reduces inference latency and input token costs for long contexts that are reused across queries, such as a document users ask several questions about.

• DRaise the temperature so the model reads the manual faster

  Incorrect

  Temperature only influences how random the output is. It does not change how much input the model processes.

Prompt caching support is a model selection criterion: for long, repeated context it cuts both latency and the input token bill.

• AA maximum response length

  Correct

  Response length parameters set the minimum or maximum number of tokens returned in the generated response.

• BA higher temperature

  Incorrect

  Temperature changes how random the token choices are. It does not cap the length of the answer.

• CA higher Top K value

  Incorrect

  Top K sets how many of the most likely candidates are considered for each next token; a higher value widens choice, not length limits.

• DA larger context window for the model

  Incorrect

  The context window is how much input the model can consider. It does not limit how long the response is.

Length parameters cap or penalize output length; randomness parameters (temperature, Top K, Top P) shape word choice.

• AA new Amazon OpenSearch Service domain for the vectors

  Incorrect

  OpenSearch Service is a strong vector store, but it is a separate service, which the retailer wants to avoid.

• BAmazon Translate to normalize the product descriptions

  Incorrect

  Amazon Translate translates text between languages. It does not store embeddings or run similarity searches.

• CThe AWS Glue Data Catalog to index the product rows

  Incorrect

  The Data Catalog stores metadata about data sources for discovery and ETL. It does not run vector similarity searches.

• DThe pgvector extension in Aurora PostgreSQL

  Correct

  Aurora PostgreSQL supports the pgvector extension to store embeddings in the database and perform efficient similarity searches.

When the data already lives in PostgreSQL, pgvector adds vector search in place; a dedicated store such as OpenSearch Service suits large search workloads.

• ASupervised fine-tuning on labeled examples

  Incorrect

  Fine-tuning adjusts the model's parameters with a training dataset.

• BIn-context learning with a few examples in the prompt

  Correct

  Few-shot prompting, also called in-context learning, puts examples in the prompt to calibrate the output; nothing in the model is trained.

• CContinued pre-training on unlabeled domain text

  Incorrect

  Continued pre-training is a training job on large amounts of unlabeled data that adapts the model to a domain.

• DDistillation into a smaller student model

  Incorrect

  Distillation fine-tunes a student model on responses from a teacher, so it is a training job.

• ERetrieval Augmented Generation over the company's documents

  Correct

  RAG adds relevant information to the prompt at query time and extends the model to the company's knowledge without retraining it.

The cost ladder runs from prompt changes and RAG (no training) through fine-tuning and distillation to continued and full pre-training (most expensive).

• ANo chunking, so that each whole document is a single chunk

  Incorrect

  With no chunking each document is one chunk, which gives up the precision of small chunks, the opposite of the goal.

• BHierarchical chunking with parent and child chunks

  Correct

  Hierarchical chunking retrieves small, precise child chunks and then replaces them with broader parent chunks, so the model gets more comprehensive context.

• CFixed-size chunking with the smallest possible chunk size

  Incorrect

  Smaller fixed chunks make matches more precise but give the model even less surrounding context.

• DDefault chunking of about 300 tokens per chunk

  Incorrect

  Default chunking splits text into chunks of about 300 tokens; it does not combine precise matching with broader context.

Hierarchical chunking separates what you search with (small child chunks) from what you give the model (larger parent chunks).

• AA larger embeddings model for the knowledge base

  Incorrect

  Embeddings improve retrieval of relevant text; they do not let the assistant call an API.

• BA lower temperature for the existing chatbot

  Incorrect

  Temperature affects how varied the generated text is, not whether the model can take actions.

• CAn AI agent that can plan steps and call tools

  Correct

  AI agents perform self-directed tasks toward a goal, and they can invoke APIs and company systems to execute tasks, which goes beyond answering from retrieved text.

• DA second knowledge base for the booking records

  Incorrect

  A knowledge base retrieves information to ground answers. Retrieval alone does not perform a booking.

RAG lets a model know things; agents let it do things, by planning and calling tools such as APIs.