Task 2.1: Explain the basic concepts of generative AI (GenAI).

• AEmbeddings, numeric vectors that can be compared for similarity

  Correct

  An embedding turns input into a vector of numbers so that objects can be compared; sentences can be compared to determine how similar their meaning is, even when the wording differs.

• BToken lists that are matched against the question word for word

  Incorrect

  A token is a unit of text a model reads, such as a word or part of a word. Matching tokens exactly still needs the same words, which is the limitation the company wants to remove.

• CA higher temperature setting so that the model considers related words

  Incorrect

  Temperature changes how random a model's generated output is. It has no role in finding which stored articles are similar to a question.

• DA model with a context window large enough to hold every article

  Incorrect

  The context window is the number of tokens a model can take into account at one time. Two million articles will not fit, and a bigger window does not provide a similarity search.

Semantic search compares meaning, and meaning is compared through embedding vectors: similar texts get vectors that are close to each other.

• ACosts stay the same, because on-demand inference is billed per hour of model runtime

  Incorrect

  Hourly billing applies to Provisioned Throughput, which is bought in model units. On-demand text models are priced per input and output token.

• BEvery request costs more, because input tokens are billed, and latency may also rise

  Correct

  On-demand text models are priced per million input tokens and per million output tokens, so a longer prompt costs more on every call. More tokens in the input can also increase inference latency.

• CCosts stay the same, because only output tokens are billed and the instructions are input

  Incorrect

  Input tokens are billed too; the pricing tables list a price per million input tokens next to the price per million output tokens.

• DCosts stay the same, because each request is billed at a flat price whatever its length

  Incorrect

  On-demand text inference is not priced per request. The price depends on how many input and output tokens are processed.

With token-based pricing, prompt length is a cost and performance lever: every token you add to a reused prompt is paid for on every request.

• AA transformer-based large language model

  Incorrect

  Large language models are transformer networks trained to understand and generate text. Gradual noising and denoising is not how they are described.

• BAn embeddings model

  Incorrect

  An embeddings model turns input into numeric vectors for comparing similarity. It does not generate new images.

• CA diffusion model

  Correct

  Diffusion models add controlled noise to data over several iterations and then reverse the process, denoising step by step to produce a new data sample.

• DA recommendation model trained on user clicks

  Incorrect

  A recommendation model suggests items to users from their past interactions. It does not create images.

Noise in, noise out, step by step: that iterative denoising is the defining mechanism of diffusion models, which many image generators use.

• AThey are used to fine-tune the foundation model so that it memorizes the policies

  Incorrect

  A knowledge base removes the need to keep training the model on your private data. The model stays as it is; the documents are retrieved at query time.

• BEach data source is turned into one embedding that represents all of its documents at once

  Incorrect

  Documents are split into many chunks first, and each chunk gets its own embedding, so retrieval can return the relevant passage.

• CThey are stored as token IDs, and questions are answered by exact token matching

  Incorrect

  The stored chunks are converted to vector embeddings so that queries and text can be compared by semantic similarity, not by exact match.

• DThey are split into chunks, and each chunk is turned into an embedding in a vector index

  Correct

  During ingestion, documents are split into manageable chunks, the chunks are converted to embeddings and written to a vector index, and a mapping to the original document is kept.

Chunking decides what a retrieval can return: the knowledge base embeds and searches chunks, then hands the best ones to the model as context.

• AIt gives agents a standard way to connect to external tools and data sources

  Correct

  MCP is a standard for exposing tools and context to models and agents. AgentCore Gateway, for example, converts APIs and Lambda functions into MCP-compatible tools that agents can discover and call.

• BIt compresses prompts so that the agents use fewer input tokens per request

  Incorrect

  MCP is about connecting agents to tools and context. Reusing repeated prompt prefixes to cut input token cost is what prompt caching does.

• CIt fine-tunes the agent's foundation model on the APIs' documentation

  Incorrect

  MCP makes tools available for an agent to call at run time. Adapting a model to your data by training it is model customization, such as fine-tuning, which is a separate process.

• DIt encrypts all traffic between the agents and the company's on-premises data center

  Incorrect

  MCP is not an encryption protocol. Private connectivity to AWS services is provided by features such as AWS PrivateLink.

MCP standardizes how an agent discovers and calls tools, so one integration works with any MCP-compatible agent instead of custom code for each pair.

• AMulti-agent collaboration among peer agents that negotiate roles

  Incorrect

  Multi-agent collaboration emphasizes peer-to-peer, emergent coordination without a central coordinator. This design has one agent in charge and a predefined sequence.

• BA workflow orchestration agent that delegates to subagents

  Correct

  Workflow orchestration agents coordinate multistep tasks: a central agent delegates work to subagents, tracks execution state and passes results to the next agent in the sequence.

• CA single retrieval-augmented agent answering from a knowledge base

  Incorrect

  A retrieval-augmented agent is a single-agent pattern that answers from retrieved data. It does not split work across other agents.

• DA single reasoning agent that works through the claim on its own

  Incorrect

  Workflow orchestration agents differ from agents that reason and act in isolation; here the work is delegated to several specialists.

A central coordinator with a predefined flow is workflow orchestration; peers that negotiate and divide work dynamically are multi-agent collaboration.

• AA unit of text, such as a word, part of a word or a punctuation mark

  Correct

  A token is a sequence of characters a model interprets or predicts as a single unit, which can be a word, part of a word such as "-ed", a punctuation mark or a common phrase.

• BExactly one complete word, so the token count always equals the word count of a prompt

  Incorrect

  Tokens can be parts of words, punctuation marks or common phrases, so the token count does not simply equal the word count.

• CA numeric vector that represents the meaning of a whole document for similarity search

  Incorrect

  A vector of numbers that represents meaning for comparison is an embedding, not a token.

• DA credential that grants an application permission to call the model

  Incorrect

  In billing, a token is a unit of text the model processes. Permission to call a model is a separate access control matter, managed with IAM.

Models read and write tokens, not words or characters, which is why prompts and responses are measured and priced in tokens.

• ATemperature sampling

  Incorrect

  Temperature shapes the probability distribution used to pick the next output token; it does not relate parts of the input to each other.

• BSelf-attention

  Correct

  The self-attention mechanism lets the model look at different parts of the sequence all at once and determine which parts are most important, instead of processing the data in order.

• CRecurrent processing, one token at a time

  Incorrect

  Earlier recurrent neural networks processed inputs sequentially; transformers process entire sequences in parallel.

• DChunking the input into fixed-size pieces

  Incorrect

  Chunking splits documents for retrieval in a knowledge base. It is not how a transformer relates words within a sequence.

Self-attention is the core of the transformer: every position weighs every other position, which also lets training run in parallel.

• AA unimodal text model

  Incorrect

  A unimodal model processes a single data type; a text-only model cannot read the photo.

• BAn embeddings model

  Incorrect

  An embeddings model turns input into vectors for comparing similarity. It does not answer questions.

• CA multimodal model

  Correct

  Multimodal models can integrate multiple data types, such as images and text, in the same request.

• DA text-to-speech model

  Incorrect

  Text-to-speech models such as those in Amazon Polly turn text into speech; they do not interpret photos.

Modality is a model selection criterion: if the input mixes images and text, the model must be multimodal.

• APre-train the model again from scratch on the same claims data to be safe

  Incorrect

  Pre-training from scratch is the most expensive step in the lifecycle and would discard the pre-trained model the team just adapted.

• BSkip testing, because fine-tuning always improves a model

  Incorrect

  Fine-tuning changes the model, and results have to be verified; AWS advises evaluating models on your own data before choosing one.

• CCollect user feedback from production before any testing

  Incorrect

  Feedback from production comes after deployment. A model should be evaluated before it reaches users.

• DEvaluate the model on use-case data against the success criteria

  Correct

  Evaluation, with programmatic metrics, human review or a judge model, shows whether the customized model meets the use case before it is deployed.

The lifecycle runs data selection, model selection, pre-training, fine-tuning, evaluation, deployment and feedback. Evaluation is the gate before deployment.

• ALong-term memory that persists across sessions

  Correct

  Without memory, agents treat each interaction as new. Long-term memory, such as AgentCore Memory provides, persists knowledge across sessions, while short-term memory covers a single multi-turn conversation.

• BA lower temperature setting for every request it sends

  Incorrect

  Temperature controls randomness in the output. It does not make the model remember anything.

• CA larger model with more parameters

  Incorrect

  Models do not recall prior prompts and previous requests unless the earlier interaction is included in the current prompt, whatever their size.

• DA guardrail with a denied topic

  Incorrect

  Denied topics block undesirable subjects. They do not store customer preferences.

Foundation models are stateless between requests; agent memory is the component that keeps context within and across sessions.