Task 3.1: Manage deployment infrastructure for ML and AI model types.

• AA real-time endpoint on a large GPU instance

  Incorrect

  Real-time inference supports payloads up to 25 MB and 60-second processing for regular responses, and the instances keep running during idle periods.

• BAsynchronous inference that scales to zero when idle

  Correct

  Asynchronous inference queues requests, accepts payloads up to 1 GB with processing times up to one hour, and can scale the instance count to zero when there are no requests to process.

• CServerless inference with the maximum memory setting

  Incorrect

  Serverless inference supports payloads up to 4 MB and processing times up to 60 seconds, far below a 400 MB clip that takes 20 minutes.

• DA nightly batch transform job over all clips uploaded that day

  Incorrect

  Batch transform suits offline processing of data that is available up front. Waiting for a nightly run would miss the one-hour target for clips uploaded in the morning.

Choose the inference option by payload size, processing time, latency target and traffic pattern. Large payloads with long processing and near real-time needs point to asynchronous inference, which can also scale to zero.

• AA separate real-time endpoint for each customer model

  Incorrect

  Two thousand single-model endpoints would each keep their own instances running, most of them idle, which multiplies cost and operational work.

• BA multi-container endpoint with one container for each customer model

  Incorrect

  Multi-container endpoints run several different containers, such as different frameworks, on one endpoint. They are not built to hold thousands of models that share one framework.

• CA batch transform job for every customer request

  Incorrect

  Batch transform is for offline processing without a persistent endpoint. Starting a job per request would add far more delay than an occasional cold start.

• DA multi-model endpoint that loads models from S3 as needed

  Correct

  Multi-model endpoints host large numbers of models that use the same framework on a shared serving container and fleet, loading models into memory as needed. They fit a mix of frequently and rarely used models when occasional cold starts are acceptable.

Many similar models with uneven traffic is the textbook case for a multi-model endpoint. Models with much higher traffic or strict latency can still get dedicated endpoints.

• AA computer vision model that needs a GPU to meet its latency target

  Incorrect

  GPUs are among the real-time inference features that serverless inference does not support.

• BA model that must reach a private database through the company VPC

  Incorrect

  VPC configuration is not supported for serverless inference, so the endpoint could not reach resources inside the VPC.

• CA document classification model that receives 50 MB payloads

  Incorrect

  Serverless inference supports payloads up to 4 MB, so 50 MB requests would need another option such as asynchronous inference.

• DAn internal tool called a few times an hour that tolerates cold starts

  Correct

  On-demand serverless inference suits workloads with idle periods between bursts that can tolerate cold starts. It scales the endpoint down to zero when there are no requests, so the tool costs nothing while idle.

• EA predictable morning burst, served with provisioned concurrency

  Correct

  Serverless inference with provisioned concurrency keeps capacity initialized for predictable bursts, so requests are served with predictable performance.

Serverless inference trades a few capabilities (GPUs, VPC configuration, large payloads) for zero idle cost and no instance management. Provisioned concurrency removes cold starts for predictable peaks.

• AAmazon Bedrock Custom Model Import

  Correct

  Custom Model Import brings models customized in other environments, such as SageMaker AI, into Bedrock. The imported model is used with on-demand throughput through InvokeModel and InvokeModelWithResponseStream.

• BA SageMaker AI real-time endpoint on GPU instances

  Incorrect

  A real-time endpoint can host the model, but the team would choose and run the instances and call a different API, which the requirements rule out.

• CAmazon Bedrock Model Distillation

  Incorrect

  Distillation trains a smaller student model from a teacher's responses. It does not bring an existing set of weights into Bedrock.

• DAn Amazon Bedrock prompt router

  Incorrect

  A prompt router sends requests between foundation models in the same model family. It cannot serve a model whose weights the team trained elsewhere.

To use the Bedrock API with a model trained outside Bedrock, import it. To manage the serving infrastructure yourself instead, host it on SageMaker AI.

• AA higher numberOfResults value for retrieval

  Incorrect

  Retrieving more chunks adds more loosely related text and more input tokens, which works against both goals.

• BHierarchical chunking on the data source

  Incorrect

  Changing the chunking strategy means re-ingesting the data, and returning parent chunks makes each result larger, not more selective.

• CA global cross-Region inference profile for generation

  Incorrect

  Global cross-Region inference changes where requests are processed and can lower price, but it does not change which chunks are retrieved.

• DA reranker model in the retrieval request

  Correct

  A reranker scores each retrieved chunk's relevance to the query and reorders the results, so the application can pass fewer but more relevant chunks to the model, which also lowers cost and latency.

Vector similarity gets candidates; a reranker orders them by relevance to the actual question. Retrieving a wider set and reranking to a small top set is a common RAG configuration.

• AAsynchronous inference with an SNS notification

  Incorrect

  Asynchronous inference queues requests and writes results to S3 later, with an optional notification, so the caller cannot get the score while waiting.

• BBatch transform run every few minutes

  Incorrect

  Batch transform is for offline processing of data available up front; running it every few minutes still cannot answer a waiting customer.

• CServerless inference with on-demand capacity

  Incorrect

  Serverless inference suits intermittent or unpredictable traffic, not steady high-volume traffic with strict latency.

• DA real-time endpoint on chosen instances

  Correct

  Real-time inference suits online inferences with low latency or high throughput needs, using a persistent, fully managed endpoint backed by the instance type you choose.

Steady, latency-critical, synchronous traffic is the home of real-time endpoints.

• AAmazon Bedrock AgentCore Runtime

  Correct

  AgentCore Runtime is framework agnostic, working with LangGraph, Strands, CrewAI and custom agents, and works with any large language model, providing a secure, serverless hosting environment.

• BBedrock Agents Classic, rebuilt in its console

  Incorrect

  Rebuilding in another service is the rewrite the team wants to avoid, and Agents Classic is no longer open to new customers.

• CA SageMaker AI batch transform job

  Incorrect

  Batch transform scores datasets offline; it does not host interactive agents.

• DAmazon Bedrock Prompt management

  Incorrect

  Prompt management stores and versions prompts; it does not host agent code.

AgentCore services work with any framework and any model, so existing agents can move to managed infrastructure without a rewrite.

• ARetrieve only, then display the returned chunks

  Incorrect

  Retrieve returns chunks without generating an answer, so the application would still need its own model call to answer the question.

• BRetrieveAndGenerate on the knowledge base

  Correct

  RetrieveAndGenerate queries the knowledge base and generates a response based on the retrieved chunks, returning citations to the original source data.

• CInvokeModel without the knowledge base

  Incorrect

  Calling the model alone gives no retrieval and no citations to policy documents.

• DStartIngestionJob for each question

  Incorrect

  Ingestion jobs sync data into the knowledge base; they do not answer queries.

For quick RAG integration with citations, RetrieveAndGenerate does retrieval and generation together; use Retrieve when you need full control of the prompt.

• ABatch inference needs Provisioned Throughput first

  Incorrect

  The documented limitation is that imported models are not supported for batch inference at all.

• BBatch jobs are limited to 1,000 documents

  Incorrect

  There is no such limit stated; the issue is that imported models are not supported for batch inference.

• CImported models cannot use Bedrock batch inference

  Correct

  Custom Model Import cannot be used with Bedrock batch inference, so the overnight run needs another approach, such as on-demand calls or hosting the model on SageMaker AI for batch transform.

• DBatch inference needs the model in another Region

  Incorrect

  Region is not the issue; imported models are excluded from batch inference.

Check feature compatibility for imported models: they work with on-demand inference but not with every Bedrock feature, such as batch inference.

• AA CPU instance with the default inference container

  Incorrect

  A CPU instance gives up the accelerators the model needs, and the default setup does not shard the model across devices.

• BA batch transform job on a single GPU instance

  Incorrect

  Batch transform is for offline processing without a persistent endpoint, and one GPU still cannot hold the weights.

• CThe training container image reused for hosting

  Incorrect

  A training image is built to run training jobs; serving a large LLM needs an inference stack such as an LMI container.

• DAn LMI container with tensor parallelism

  Correct

  Large model inference (LMI) containers are specialized containers for LLM inference that support features such as tensor parallelism, which spreads the model across the instance's GPUs, and continuous batching.

When a model is too large for one accelerator, use a serving container that shards it across the accelerators of the instance.