Foundation Model (FM) Integration, Data Management, and Compliance — AWS Certified Generative AI Developer AIP-C01 practice questions

An auditor asks a GenAI team to demonstrate which source documents contributed to a specific chatbot answer and when those documents were last updated. Which capability MOST directly supports this requirement?

1. Switching the embeddings model to one with a higher maximum context window
2. Data lineage and audit logging that tracks document ingestion, versioning, and which chunks were retrieved and cited for a given response
3. Increasing the foundation model's temperature so answers vary less between requests
4. Enabling provisioned throughput for the model to guarantee consistent response latency

Answer: B — Data lineage and audit logging that tracks document ingestion, versioning, and which chunks were retrieved and cited for a given response

Right — demonstrating which documents contributed to an answer and their update history requires explicit data lineage and audit logging across ingestion, versioning, and retrieval, which is the controlling capability for this compliance request. Temperature affects output variability, not traceability, and lowering it does not create any lineage record. Provisioned throughput addresses latency and capacity guarantees, an operational concern unrelated to auditability. A higher-context embeddings model is a true capability of some models but does not by itself produce a lineage trail of which documents were used and when they were updated.

A developer notices that a model performs inconsistently on a multi-step arithmetic word problem. Which prompting technique is MOST likely to improve accuracy without changing the model or adding external tools?

1. Lowering the temperature to 0 and removing all instructions from the prompt
2. Increasing the top-k parameter to allow more diverse token choices
3. Switching to zero-shot prompting to reduce prompt length and latency
4. Instructing the model to reason through the problem step by step before giving a final answer (chain-of-thought prompting)

Answer: D — Instructing the model to reason through the problem step by step before giving a final answer (chain-of-thought prompting)

Right — chain-of-thought prompting is well suited to improving accuracy on multi-step reasoning by encouraging the model to work through intermediate steps. Lowering temperature to 0 can reduce randomness, but removing instructions is a half-right idea that strips away the guidance that actually helps reasoning. Increasing top-k pushes toward more diverse (less deterministic) output, which is the opposite of what stabilizes a reasoning task. Switching to zero-shot removes structure that could help, prioritizing latency over the actual accuracy problem described.

Which AWS service provides a curated hub of pre-trained foundation and task-specific models, along with example notebooks, that a team can deploy directly to a SageMaker endpoint for fine-tuning or inference?

1. Amazon SageMaker Canvas
2. AWS Marketplace for Machine Learning
3. SageMaker JumpStart
4. Amazon Bedrock Knowledge Bases

Answer: C — SageMaker JumpStart

Right — JumpStart is the model hub inside SageMaker Studio that packages pre-trained models with deployable notebooks for fine-tuning and inference. Knowledge Bases is a real Bedrock capability, but it manages RAG retrieval, not a general model catalog with notebooks. SageMaker Canvas is a true AWS service, but it is a no-code ML/BI tool, not a model hub. AWS Marketplace does list ML models for purchase, which makes it tempting, but it is a general software marketplace rather than the SageMaker-integrated hub with example notebooks.

Source documents in an Amazon S3 bucket that backs a Bedrock Knowledge Base are updated weekly. Which approach ensures the knowledge base reflects those updates without re-embedding the entire corpus from scratch every time?

1. Manually delete and recreate the entire knowledge base every week
2. Rely on the foundation model to automatically detect stale context at query time and self-correct
3. Disable chunking so updates apply at the whole-document level instead of the chunk level
4. Configure an incremental data source sync that detects and re-embeds only new or changed documents since the last sync

Answer: D — Configure an incremental data source sync that detects and re-embeds only new or changed documents since the last sync

Right — an incremental sync that processes only new or changed documents is the efficient, standard approach to keeping a knowledge base current without full re-embedding. Deleting and recreating the entire knowledge base weekly is the direction-reversed, brute-force approach that the question explicitly asks to avoid. Foundation models have no mechanism to detect staleness of retrieved context or self-correct outdated knowledge base content; that is a common misconception about how RAG grounding works. Disabling chunking does not address sync efficiency and would actually harm retrieval granularity, an off-target change.

A developer wants a model's output to always conform to a strict JSON schema so a downstream service can parse it reliably. Beyond lowering temperature, which technique MOST directly increases the likelihood of schema-conformant output?

1. Increasing the maximum context window available to the model
2. Switching from a chat-style API call to a text-completion-style API call
3. Reducing the number of few-shot examples so the model has more room to improvise
4. Providing an explicit output format instruction with a schema definition or example JSON structure in the prompt

Answer: D — Providing an explicit output format instruction with a schema definition or example JSON structure in the prompt

Right — explicitly specifying the target schema or a JSON example in the prompt is the most direct lever for shaping structured output, addressing the controlling requirement. A larger context window is a true capability but does not itself constrain output format. Switching API styles is an off-target configuration change that does not address formatting compliance. Reducing few-shot examples removes exactly the kind of demonstration that helps the model match a target format, so it works against the goal.

A team needs to generate marketing images from text prompts as part of a content pipeline running entirely within AWS. Which category of foundation model on Bedrock should they select?

1. A text-to-image diffusion model such as Stable Diffusion or Amazon Titan Image Generator
2. Amazon Titan Text Embeddings, since embedding models can be adapted to generate pixel data
3. A Cohere Command text-generation model, since it can describe images in natural language
4. Amazon Titan Text Premier, because it has the largest context window in the Titan family

Answer: A — A text-to-image diffusion model such as Stable Diffusion or Amazon Titan Image Generator

Right — image generation requires a diffusion-based text-to-image model; that modality match is the controlling factor. Embeddings models is a common misconception distractor: embeddings produce vector representations for similarity, not pixel output. The Cohere text model can describe an image in words but cannot render one, which is a right-concept-wrong-modality trap. Titan Text Premier's context window is a true fact about a different model that is irrelevant to an image-generation requirement.

A financial services company must keep all inference traffic within a predictable, reserved throughput capacity to meet a strict SLA during peak trading hours, and is willing to commit to a fixed monthly cost. Which Bedrock capability BEST fits this requirement?

1. Using standard on-demand invocation with retry logic to absorb throttling
2. Enabling cross-region inference profiles to route requests to whichever region has capacity
3. Requesting a service quota increase for on-demand tokens per minute
4. Purchasing Provisioned Throughput for the chosen model

Answer: D — Purchasing Provisioned Throughput for the chosen model

Right — Provisioned Throughput reserves dedicated model capacity for a fixed commitment, which directly satisfies a guaranteed-throughput SLA. On-demand with retries is a half-right approach: it can help with transient throttling but offers no dedicated capacity guarantee. Cross-region inference profiles improve availability by routing across regions, but that is a different mechanism than a reserved capacity commitment. A quota increase raises the ceiling on shared on-demand capacity but still does not reserve dedicated throughput the way Provisioned Throughput does.

Which chunking approach groups sentences or passages based on measured semantic similarity rather than a fixed character or token count?

1. Fixed-size chunking
2. Recursive character chunking
3. Sliding-window token chunking with a constant stride
4. Semantic chunking

Answer: D — Semantic chunking

Right — semantic chunking explicitly uses similarity measures between sentences or passages to decide chunk boundaries, matching the question's description. Fixed-size chunking is the direction-reversed approach, splitting purely by count regardless of meaning. Recursive character chunking splits along structural delimiters (like paragraphs) before falling back to character counts, which is structure-based rather than similarity-based. Sliding-window token chunking with constant stride is another fixed-count method, just with overlap, and does not consider semantic similarity.

A developer ingests scanned PDF contracts into a RAG pipeline but retrieval quality is poor because the text extracted from the PDFs is garbled. Which step in the pipeline is the MOST likely root cause?

1. The document ingestion/text-extraction step failed to properly OCR or parse the scanned images before chunking and embedding
2. The embeddings model chosen does not support cosine similarity as a distance metric
3. The vector store's index type does not support approximate nearest neighbor search
4. The chunk size was configured too small for the embeddings model's minimum input length

Answer: A — The document ingestion/text-extraction step failed to properly OCR or parse the scanned images before chunking and embedding

Right — garbled text pointing back to the source content strongly indicates a failure earlier in the pipeline, at OCR/text extraction from the scanned images, which is the root cause of downstream embedding and retrieval quality problems. Most embeddings models do support cosine similarity, and even if a different metric were used, that would affect ranking, not cause garbled extracted text. Approximate nearest neighbor support is a vector-store indexing capability unrelated to the quality of the text that was extracted. An overly small chunk size could hurt retrieval granularity, but it would not explain garbled character-level text content, which points specifically to the extraction step.

A developer provides a foundation model with a task description and three worked examples of input/output pairs before asking it to complete a new, similar task. Which prompting technique is being used?

1. Few-shot prompting
2. Zero-shot prompting
3. Chain-of-thought prompting
4. Retrieval-augmented generation

Answer: A — Few-shot prompting

Right — including a small number of worked examples alongside the instruction is the defining characteristic of few-shot prompting. Zero-shot prompting is the direction-reversed case with no examples at all. Chain-of-thought prompting is a different technique focused on eliciting step-by-step reasoning, not on providing input/output examples. Retrieval-augmented generation is an architecture for grounding responses in external documents, unrelated to supplying examples directly in the prompt.

What is Stable Diffusion used for on Amazon Bedrock?

Stability AI's Stable Diffusion models generate and edit images from text prompts, supporting use cases like creative content generation, image inpainting, and image-to-image transformation.

What does the temperature parameter control during inference?

Temperature controls the randomness of token selection; lower values (near 0) produce more deterministic, focused output, while higher values increase diversity and creativity at the cost of predictability.

What vector engine does Amazon OpenSearch Serverless provide for GenAI workloads?

It provides a fully managed, auto-scaling vector search engine that stores embeddings and performs low-latency k-nearest-neighbor (kNN) similarity search without the operator needing to provision or manage clusters.

What is the AWS Certified Generative AI Developer - Professional exam?

It is a professional-level AWS certification that validates a candidate's ability to design, build, deploy, secure, and optimize generative AI applications on AWS using services such as Amazon Bedrock and Amazon SageMaker. It targets experienced GenAI developers who work with foundation models in production.