AIF-C01 — frequently asked questions

Comprehend extracts entities, sentiment, and key phrases from text only; it cannot process audio recordings, so transcription with Transcribe must happen first.

Transcribe focuses on speech-to-text; deeper NLP extraction such as sentiment or entities is typically done afterward with a text service like Comprehend.

An endpoint is only the hosting target; it has no foundation model catalog or prebuilt containers. JumpStart is the feature that supplies the FM and deploys it to that endpoint quickly.

Yes. Models deployed from JumpStart run on SageMaker endpoints in your account, and you can apply VPC configuration and network isolation so traffic never traverses the public internet.

Rekognition analyzes images and video for objects, faces, and text, not free-form review sentiment. Written reviews require NLP or a foundation model.

Yes. You can prompt a foundation model in Bedrock to classify review sentiment, or customize a model for your domain.

Model access grants unlock third-party base foundation models for an account. A custom model you fine-tuned already exists in your account, but it still needs Provisioned Throughput before it can serve inference.

No. Bedrock custom models have no on-demand option; you must purchase Provisioned Throughput for that model, which provisions dedicated model units for your workload.

Few-shot examples only steer behaviour within the context window; they cannot add vocabulary or domain knowledge the FM lacks. The stem already says multiple prompt engineering attempts failed.

No. Removing complex terms strips the precise vocabulary the chatbot must interpret, degrading answers for a research audience instead of teaching the model the domain.

Clarify provides bias metrics and explanations, and Model Monitor can use those analyses, but Clarify alone is not the ongoing production drift-monitoring feature for a deployed model.

Model Monitor supports bias drift monitoring as well as data quality and model quality monitoring; Clarify supplies the bias analysis used by those jobs.

CloudWatch can monitor Comprehend endpoint metrics, but the question asks for the built-in report of endpoints unused for over 15 days, which Trusted Advisor provides.

Yes, the check identifies Comprehend endpoints not used for real-time inference for more than 15 days and refreshes results automatically.

No, SSE-S3 uses S3-managed keys. You do not need KMS permissions, but you do need S3 permissions like s3:GetObject.

Never. Public access violates security best practices and compliance requirements, especially for potential sensitive data.

Context window refers to the maximum number of tokens (input + output) a model can process at once. Model size refers to the number of parameters in the model, which affects complexity and performance, not necessarily input length.

Not necessarily. A larger context window allows for longer inputs but does not guarantee higher intelligence or accuracy. You must balance context window size with other factors like reasoning capability and cost.

Data Wrangler is for interactive data preparation and exploration. Feature Store is for storing, versioning, and sharing prepared features across teams and projects for consistent model training and inference.

It prevents silos by allowing teams to discover, use, and contribute to a common set of features. This reduces redundancy and ensures all models use consistent definitions for the same variables.

Amazon Q Developer is a generative AI assistant designed to improve developer productivity by helping them write, understand, and secure code faster.

No, Amazon Q Developer primarily interacts through natural language text prompts within integrated development environments (IDEs), not voice commands.

An Internet Gateway routes traffic to the public internet. Since the requirement prohibits all internet traffic, an IG violates the compliance constraint.

Yes, AWS PrivateLink supports many AWS services including Bedrock, allowing private connectivity from your VPC to these services.

Reinforcement learning is for sequential decision-making in dynamic environments with reward signals. It is unsuitable for static, deterministic calculations like probability ratios.

Yes. While ML models might take longer to develop, the primary overhead in AWS exams usually refers to runtime compute, storage, and management complexity. Simple code has negligible overhead in both.

Training time measures how long it takes to build or fine-tune a model using a dataset. Response time measures how long it takes for the trained model to generate a prediction when processing a single input request.

CSAT is a subjective business metric reflecting user happiness. It does not technically measure the computational speed or latency of the AI model itself.

No, supervised learning strictly requires labeled training data where inputs are paired with known correct outputs. Without labels, the model has no target to learn from.

It refers to clustering, where the algorithm groups similar data points together. These groups are then interpreted as 'tiers' or segments by business analysts.

No. Provisioned throughput is required for inference capacity after the model is fine-tuned. It is not part of the training/fine-tuning process itself.

Data should be provided as labeled examples with 'prompt' (input) and 'completion' (desired output) fields, typically in JSONL or CSV formats supported by Bedrock.

Anomaly detection uses machine learning to establish a baseline of normal network behavior. It then flags any activities, such as access from a specific IP address, that significantly deviate from this baseline as potential threats.

Natural Language Processing (NLP) is designed for text and speech data. Since IP addresses and network logs are numerical and structural data, NLP techniques like Named Entity Recognition are not applicable for analyzing traffic patterns.

No, forecasting is primarily predictive analytics. While generative models can analyze data, the act of predicting future numerical trends based on historical patterns is distinct from generating new content like images or text.

It uses learned patterns from training data to create entirely new, previously non-existent instances (like images) that resemble the training distribution, rather than just retrieving or classifying existing ones.

Fine-tuning is more resource-intensive and time-consuming. Prompt engineering is the fastest, most cost-effective way to adjust tone and style for specific use cases without modifying the model weights.

While higher temperature increases creativity, it reduces consistency. For a company-specific tone, consistency is usually prioritized, so prompt engineering offers better control than randomization.

Few-shot prompting is a technique where you provide the LLM with a few examples of the task (input-output pairs) in the prompt before asking it to perform the task on new data.

While zero-shot can work, few-shot provides clearer guidance and consistency by showing the model exactly what kind of labels and formatting are expected, reducing errors.

No. AWS Data Exchange is for subscribing to third-party data products like market or weather data. Compliance reports are handled by AWS Artifact.

It is a feature in AWS Artifact that sends email alerts to users when new compliance reports, agreements, or certifications are published.

Prompt injection is a technique where attackers manipulate an LLM by inserting malicious instructions into the input, causing the model to perform unintended actions or leak data.

No. Increasing temperature makes outputs more random and creative, which can actually make the model less predictable and potentially more susceptible to following erratic injected instructions.

RAG does not require modifying the model weights or performing expensive training runs. It simply retrieves relevant documents and adds them to the prompt, incurring only storage and retrieval costs.

Prompt engineering can guide tone and format, but it cannot provide the model with new, proprietary information or update its knowledge base like RAG does.