Task 1.1: Analyze requirements and design GenAI solutions.

• ACreate an Amazon Bedrock managed knowledge base with Confluence and SharePoint connectors and ACL awareness, and pass each user's identity

  Correct

  A managed knowledge base provides the Confluence and SharePoint connectors, crawls document permissions when ACL awareness is enabled, filters results by the user context supplied at query time, and returns citations. Bedrock manages the data store, so no vector database has to be operated.

• BFine-tune a foundation model on an export of the guidelines every week and serve it through an on-demand custom model deployment

  Incorrect

  Fine-tuning bakes the content into model weights, so answers lag behind changes made between training runs, cannot cite a source page, and cannot enforce per-team visibility.

• CCreate a customer-managed knowledge base on Amazon OpenSearch Serverless and copy the pages into Amazon S3 with a nightly export job

  Incorrect

  A customer-managed knowledge base means provisioning and managing the vector store, and the third-party connectors and document-level permission filtering are offered only for managed knowledge bases. A nightly copy to S3 also drops the source permissions.

• DPlace all guideline pages in the system prompt of every request and instruct the model to quote the title of each page it relied on in the answer

  Incorrect

  Several systems' worth of pages do not fit in a context window, every request would pay for all of those tokens, and a prompt instruction cannot enforce which employee may see which page.

Frequently changing internal content with citations is a retrieval problem, not a training problem. When the sources are third-party systems with their own permissions and nobody wants to run a vector database, the managed knowledge base is the fit: native connectors, ACL-aware retrieval and a Bedrock-managed store. The application must still authenticate the user, because ACL awareness filters on the identity it is given.

• ARun a model distillation job that uses a larger teacher model to generate responses for the questions

  Incorrect

  Distillation transfers a larger model's behavior to a smaller one by generating teacher responses. The goal here is the human-written house style that already exists as labeled data, not a teacher's answers.

• BRun a reinforcement fine-tuning job with a Lambda reward function that scores how polite each response is

  Incorrect

  Reinforcement fine-tuning suits cases where you can score responses but lack labeled outputs. Here labeled outputs exist, and a politeness score does not describe the required structure and phrasing.

• CBuild a knowledge base from the 6,000 pairs and retrieve the five most similar pairs into each prompt at query time

  Incorrect

  Retrieval supplies knowledge at query time, but the problem is consistent output form, not missing facts. Retrieved examples still sit in every prompt and keep the requests large.

• DFine-tune a supported base model with the 6,000 labeled pairs and compare it with the prompt-only baseline

  Correct

  Supervised fine-tuning learns from labeled input and output pairs to produce a specific kind of output, which is what a consistent style needs. It also removes the long few-shot block from every request.

The AWS guidance is to start with prompting and retrieval, and to customize when a specific task needs better performance and labeled data exists. A stable style that prompting cannot reach reliably, plus thousands of examples of the exact output, is the classic case for supervised fine-tuning. Distillation and reinforcement fine-tuning answer different data situations.

• AReview each workload in the AWS Well-Architected Tool with the Generative AI Lens and a shared custom lens

  Correct

  The Well-Architected Tool lens catalog includes the Generative AI lens, custom lenses add organization-specific questions and can be shared, and each workload review produces improvement items that can be tracked over time.

• BRun AWS Trusted Advisor checks on each account before launch and attach the report to the approval ticket

  Incorrect

  Trusted Advisor checks accounts for cost, security and service limit issues. It is not a structured review of a generative AI workload's design.

• CDeploy an AWS Config conformance pack to every account and treat its compliance score as the design review result

  Incorrect

  Conformance packs evaluate resource configurations against rules. They do not cover architectural questions such as model choice, prompt management or evaluation practices.

• DPublish an approved generative AI architecture as an AWS Service Catalog product that every business unit must launch

  Incorrect

  Service Catalog standardizes what gets provisioned, which is useful, but it does not review a design against best practices or track an improvement plan for each workload.

Standardizing how designs are reviewed is what the Well-Architected Tool does: a lens supplies the questions and best practices, a custom lens adds the company's own, and every review records risks and improvement plans in one place. The Generative AI lens is available in the tool's lens catalog.

• AAn Amazon Bedrock knowledge base with a vector store for the policy PDFs, queried with RetrieveAndGenerate

  Correct

  Unstructured PDFs are parsed, chunked and embedded into a vector store, and RetrieveAndGenerate returns a generated answer with citations to the source chunks.

• BA model fine-tuned on last year's revenue tables so it can answer revenue questions directly

  Incorrect

  A fine-tuned model cannot see hourly updates and would produce figures from its weights rather than computing them in the database.

• CA nightly export of the Redshift tables to CSV files that are ingested into a vector knowledge base

  Incorrect

  This copies finance data into another store, which finance forbids, is up to a day stale, and leaves arithmetic to the model reading text chunks.

• DAn Amazon Bedrock knowledge base connected to the Redshift tables as a structured data store

  Correct

  A knowledge base connected to a structured data store converts natural language questions into SQL that runs on Redshift, so the figures come from the database and the data is not converted or copied into a vector store.

• EAn Amazon Bedrock Data Automation project that turns the policies into JSON that is added to every prompt

  Incorrect

  Data Automation is the right tool to extract structured fields from documents, but sending the whole policy set with every request wastes tokens and does not produce citations to a source document.

The question mixes two kinds of data. Tabular data that must stay in place and be calculated exactly is served by a knowledge base connected to a structured data store, which turns the question into SQL. Unstructured documents that need citations are served by a vector knowledge base with RetrieveAndGenerate.

• AFine-tune a model on the catalog export once a month and serve it on demand

  Incorrect

  Supervised fine-tuning learns a task from labeled examples, which the team does not have, and a monthly model is already out of date for the products that change every day. It also cannot point to the product page behind a fact.

• BIndex the export in a knowledge base, sync it after each export, and answer with citations

  Correct

  A knowledge base retrieves only the passages relevant to each question and returns citations. Syncing is incremental, so each nightly sync processes just the added, changed or deleted products.

• CSend the whole catalog export with every request to a model with a large context window

  Incorrect

  Placing reference text in the prompt suits small documents. A 30,000-product catalog in every one of 2,000 requests an hour is a very large and costly prompt.

• DRun a distillation job each night so that a smaller model learns the latest catalog

  Incorrect

  Distillation fine-tunes a smaller student model on a teacher model's responses to cut inference cost for a use case. It adds a training run every night and still gives no source page to cite.

Facts that change daily and must be cited belong in a retrieval layer, not in model weights. AWS guidance puts prompt engineering and RAG before customization, and a synced knowledge base keeps both freshness and per-request cost under control.

• ATrain the model with reinforcement fine-tuning against a brand-tone reward function

  Incorrect

  Reinforcement fine-tuning trains a model against reward functions. It is a training job that would have to be repeated whenever the tone guidance changes, and AWS guidance places customization after prompt engineering and RAG.

• BFine-tune a model on the 12 approved descriptions with supervised fine-tuning

  Incorrect

  Supervised fine-tuning learns from a labeled training dataset. Twelve examples are a very small dataset, and every revision of the guidance would need a new training run.

• CBuild a knowledge base of past descriptions and retrieve similar ones for each request

  Incorrect

  RAG supplies external information to a response. Tone and structure are requirements on the output that the prompt can state directly.

• DWrite clear tone and structure instructions in the prompt with two approved examples

  Correct

  AWS guidance is to start with prompt engineering, the least resource-intensive option. Clear instructions plus a few examples set tone and structure, and the prompt can be edited the day the guidance changes.

Prompt engineering comes first: it is fast, cheap and easy to change. Move to RAG for missing knowledge and to customization only when prompting cannot reach the goal and enough training data exists.

• AA knowledge base for the manual plus a reservations API tool

  Correct

  Retrieval suits static documents, and a tool call fetches live, per-customer data at request time.

• BA prompt that includes the full manual and the customer's booking history

  Incorrect

  A 400-page manual in every prompt is costly and slow, and history does not give the live status.

• CA fine-tuned model trained on the manual and on past bookings

  Incorrect

  Training cannot know today's booking status and does not cite the manual.

• DA knowledge base that indexes nightly exports of the reservations data

  Incorrect

  Nightly exports are stale for status that changes every minute, and they mix personal data into a shared index.

Combine retrieval for documents with tools for live, transactional data. Each source is used where it is accurate.

• AAmazon Bedrock Agents Classic with action groups for each internal API

  Incorrect

  Bedrock Agents Classic served earlier customers, but it is no longer open to new customers starting July 30, 2026.

• BAWS Step Functions with one Task state for each internal API call

  Incorrect

  Step Functions suits workflows whose steps are defined in advance. Here the model chooses the calls, and a tool layer and cross-session memory would still have to be built.

• CAmazon Bedrock AgentCore, using Gateway for tools and Memory for context

  Correct

  Gateway turns existing APIs into MCP-compatible tools that agents reach through Gateway endpoints, and Memory keeps short-term and cross-session context, both as managed services.

• DAn open-source agent framework on Amazon ECS with AWS Fargate

  Incorrect

  Fargate removes server management, but the team would still build and operate its own memory store and MCP tool layer.

Pick the agent platform by what the team must not build: a managed tool gateway and managed memory remove most of the undifferentiated work, and lifecycle status rules out services closed to new customers.

• AAmazon Bedrock batch inference with the filings in Amazon S3

  Correct

  Batch inference processes many prompts asynchronously from S3 files, which suits large overnight workloads without real-time needs.

• BA streaming ConverseStream call for each filing

  Incorrect

  Streaming improves time to first token for interactive users, which this workload does not need.

• COne synchronous Converse call per filing from a single script

  Incorrect

  This works but uses on-demand pricing and must manage throttling and retries for 200,000 calls.

• DProvisioned Throughput with a six-month commitment

  Incorrect

  A long commitment pays for capacity around the clock for a nightly job.

Match the inference mode to latency needs. Large offline workloads are a natural fit for batch inference.