Task 1.3: Apply GenAI concepts and techniques.

• ARaise the temperature to vary the wording

  Incorrect

  A higher temperature makes responses more random. That helps with creative variety, but it works against the consistent labels needed here.

• BName the three allowed categories and ask for exactly one of them

  Correct

  Models work best with simple, clear and complete instructions. Naming the choices explicitly turns a vague request into a defined classification task.

• CAdd the company's full product history as background before the email

  Incorrect

  Extra context that is unrelated to the task makes the prompt longer and costlier without clarifying what output is expected.

• DAsk for only the category word in the response, with no explanation

  Correct

  Output indicators that specify the form and length of the response keep the model from adding free-form text the spreadsheet cannot use.

• EAsk the model to explain its reasoning in detail before giving a label

  Incorrect

  Step-by-step reasoning can help with complex problems, but it makes the response longer and adds exactly the free-form text that breaks the routing.

Inconsistent output usually means an ambiguous prompt. Telling the model exactly which choices exist and exactly what the answer should look like is the first and cheapest fix, before any change of model or customization.

• AIncrease the maximum output length setting for the request

  Incorrect

  The maximum output setting limits how many tokens the model may generate. It does not let a larger input fit into the context window.

• BSummarize it in sections that each fit the context window, then combine them

  Correct

  The context window limits how many tokens a model can take in at once. Chunking breaks large documents into pieces small enough to fit, so the work can be done in parts.

• CFine-tune the model on the agreement so that it can summarize the text from memory

  Incorrect

  Fine-tuning adapts a model's behavior with training examples; it is slow and costly, and it is not a reliable way to make a model hold one specific long document for a summary.

• DLower the temperature so the model processes the text more carefully

  Incorrect

  Temperature changes how random the generated tokens are. It has no effect on how much input the model can accept.

Every model has a context window, a limit on the tokens it can consider at once. When a document exceeds it, the input must be reduced, typically by chunking the document and working section by section, or by retrieving only the relevant parts. Context window size is also a criterion when choosing a model for long-document work.

• ARetrieval augmented generation that retrieves current guideline passages per question

  Correct

  RAG references an authoritative knowledge base outside the model's training data before generating a response, so updated guidelines are used as soon as they are indexed, without retraining the model.

• BFine-tune a foundation model on the guidelines every time they are revised

  Incorrect

  Fine-tuning changes the model's weights through a training job. Doing that every week is slow and costly, and for injecting knowledge AWS recommends RAG as the better starting point.

• CWrite a detailed system prompt that summarizes the guidelines

  Incorrect

  A prompt cannot hold thousands of pages, and a summary loses the details employees need. Prompting alone fits tasks that rely on the model's general knowledge.

• DContinue pretraining a foundation model on the full guideline archive

  Incorrect

  Continued pretraining adapts a model to domain language using large amounts of data. It is the costliest option and still leaves the model out of date after the next weekly revision.

When answers must rest on specific, proprietary and frequently changing documents, RAG is the standard choice: it grounds the response in the retrieved source and stays current without touching the model. It also lets the system cite the passage it used.

• AFine-tune a model on the 5,000 descriptions the copy team approved

  Correct

  Fine-tuning suits teaching a model a specific style, format or niche terminology that is hard to reproduce through prompting alone, and the approved descriptions are the high-quality examples it needs.

• BAdd a retrieval layer that pulls facts from the product database at request time

  Incorrect

  Retrieval brings in facts the model lacks, but the facts are already available. The gap is style and structure, which retrieval does not teach.

• CPretrain a new foundation model from scratch on the brand's content

  Incorrect

  Building a foundation model is the most expensive option and is reserved for cases where no available model meets the need. 5,000 examples are far too few for pretraining.

• DRaise the temperature so the descriptions sound more creative

  Incorrect

  Higher temperature increases randomness. It makes the voice and the seven-part structure less consistent, not more.

Start with prompt engineering and add retrieval when the model lacks facts. When the model has the facts but cannot reliably produce a particular style or format, fine-tuning on high-quality examples is the right next step, after weighing its cost against the expected gain.

• AKeep the per-word estimate, because Bedrock converts words to tokens one for one

  Incorrect

  Tokens do not map one to one to words; a single word can be split into several tokens, and punctuation counts too.

• BEstimate per request, because Bedrock charges a flat fee for each API call

  Incorrect

  On-demand model inference is priced by input and output tokens, so long requests cost more than short ones.

• CEstimate only the input, because output tokens are free on Bedrock

  Incorrect

  Bedrock prices output tokens separately from input tokens, so the length of the generated summaries matters to cost.

• DEstimate in tokens, which can be words, word fragments or punctuation

  Correct

  A token is a sequence of characters the model treats as one unit, such as a word, part of a word or a punctuation mark. Bedrock's on-demand prices are stated per million input and output tokens.

Tokens are the unit in which models read, generate, are limited and are billed. Because tokens do not map one to one to words, cost estimates and capacity plans should be built in tokens for both the input and the output.

• AMove to a model with a larger context window and accept any input size

  Incorrect

  A larger window only moves the limit; accepting any input size still lets attackers flood the model, and AWS's mitigation is to limit the size of the input.

• BRaise the maximum output tokens so the model can answer in full

  Incorrect

  Output length parameters limit the response, not how much input the model must hold, so they do nothing about overflow from long inputs.

• CLimit input size and keep the essential information first

  Correct

  AWS's security guidance on context window overflow recommends input management: limit the size of input, prioritize essential information and sanitize excessive content.

• DRepeat the full original instructions inside every user message

  Incorrect

  Repeating long instructions adds text to every turn, so the window fills faster, and it does nothing to stop deliberate flooding.

• EAlert when a conversation nears the context window's capacity

  Correct

  AWS also recommends real-time monitoring that triggers alerts when the context window is nearing capacity, so overflow can be managed before instructions are lost.

Models can only consider a limited amount of information at once. When inputs exceed that limit, earlier instructions can be lost, so input size must be managed and monitored rather than simply allowed to grow.

• AFine-tuning a model on all patient records so it answers from memory

  Incorrect

  Training sensitive data into a model makes it hard to control who can draw it out; retrieval keeps access controls in force.

• BPasting full patient records into every prompt for context

  Incorrect

  AWS advises not to include sensitive data in prompts beyond what the task needs.

• CRetrieval with strong authorization, not fine-tuning

  Correct

  AWS's guidance on isolating sensitive data recommends using RAG with strong authentication and authorization to augment model data over fine-tuning, alongside data minimization.

• DRemoving all access controls

  Incorrect

  Removing controls exposes records to people without a right to them.

How a model is adapted affects data security. Keeping sensitive data in governed sources and retrieving it with authorization limits exposure better than embedding it in a model.

• ALower the temperature so the model favors its most likely wording

  Incorrect

  A lower temperature steepens the probability distribution and makes responses more deterministic, which narrows variety further.

• BRaise the temperature to allow less likely word choices

  Correct

  A higher temperature increases the likelihood of lower-probability outputs and flattens the distribution, which leads to more varied responses.

• CLower Top P so the model considers fewer candidate words

  Incorrect

  A lower Top P considers a smaller share of the probability distribution, cutting less likely words and so reducing variety.

• DRaise the maximum response length so each run writes more

  Incorrect

  Response length only sets how many tokens may be generated; longer output breaks the short format without making the word choices more varied.

Inference parameters let the same model serve different needs. For brainstorming, more randomness yields more diverse options to choose from, while length settings and the prompt keep the output on format.

• ANothing else, because fine-tuning is included in inference pricing

  Incorrect

  Training and storage are charged separately from inference.

• BTraining charges per token processed, plus monthly storage

  Correct

  Bedrock charges for model training based on the number of tokens processed (tokens in the training data times epochs) and for model storage per month per model.

• COnly a one-time license fee for the base model that is being tuned

  Incorrect

  Bedrock's documented customization charges are based on training tokens and monthly storage, not a license fee.

• DOnly the cost of labeling, because training itself is free

  Incorrect

  Training is not free; tokens processed during training are charged.

Model adaptation choices have different cost profiles. Fine-tuning adds training and storage charges on top of usage, which belongs in the business case alongside the expected improvement.

• AA zero-shot prompt without the ten long examples

  Correct

  AWS's cost guidance suggests experimenting with zero-shot prompting for common knowledge tasks and keeping prompts as short as possible while meeting performance requirements.

• BAdding ten more examples to improve quality further

  Incorrect

  More examples add tokens and cost when quality is already fine.

• CFine-tuning a model first

  Incorrect

  Customization costs more than first testing a leaner prompt.

• DDoubling the maximum output length for each summary

  Incorrect

  Longer outputs raise cost without addressing the prompt overhead.

Prompt design affects cost as well as quality. When a task needs only general skills, a lean prompt can deliver the same results for fewer tokens.