AWS AI Practitioner AIF-C01 Field Notes
Use these field notes as a decision-focused review for AIF-C01, then practice with 500+ original questions. The notes follow the current public AWS blueprint: AI and ML fundamentals 20%, generative AI fundamentals 24%, foundation-model applications 28%, responsible AI 14%, and security, compliance, and governance 14%.
Short answer
The correct answer is B. Use Amazon Bedrock Knowledge Bases to retrieve relevant passages and ground the response.
Amazon Bedrock Knowledge Bases provides a managed retrieval-augmented generation pattern. The manuals are ingested, divided into retrievable chunks, represented with embeddings, and stored in a supported vector store. At request time, the system retrieves passages relevant to the question and supplies them as context to the foundation model. This keeps answers grounded in the current manuals and supports source attribution without changing the model weights. Updating the knowledge source is operationally simpler than repeatedly fine-tuning the model. Fine-tuning is better when the goal is to change stable behavior, style, or task performance—not to inject facts that change every week. Temperature controls output variability and cannot give the model access to missing private information. Amazon Comprehend is useful for managed NLP tasks such as entities, sentiment, and classification, but a custom entity recognizer does not implement grounded question answering. The exam reflex is to separate knowledge from behavior: use retrieval for changing private facts, prompt engineering for instructions and format, and fine-tuning for durable behavior that examples can teach.
The Question
A company needs a customer-support assistant to answer from private product manuals that change every week. Answers should cite the relevant source passages, and the company does not want to retrain a foundation model whenever a manual changes. Which approach best fits?
Why B is correct
Amazon Bedrock Knowledge Bases provides a managed retrieval-augmented generation pattern. The manuals are ingested, divided into retrievable chunks, represented with embeddings, and stored in a supported vector store. At request time, the system retrieves passages relevant to the question and supplies them as context to the foundation model. This keeps answers grounded in the current manuals and supports source attribution without changing the model weights. Updating the knowledge source is operationally simpler than repeatedly fine-tuning the model. Fine-tuning is better when the goal is to change stable behavior, style, or task performance—not to inject facts that change every week. Temperature controls output variability and cannot give the model access to missing private information. Amazon Comprehend is useful for managed NLP tasks such as entities, sentiment, and classification, but a custom entity recognizer does not implement grounded question answering. The exam reflex is to separate knowledge from behavior: use retrieval for changing private facts, prompt engineering for instructions and format, and fine-tuning for durable behavior that examples can teach.
Why the other options are wrong
Weekly fine-tuning is expensive and operationally heavy for facts that change frequently. Retrieval keeps the source content separate from the model weights and makes updates easier.
Temperature changes randomness. It neither retrieves private manuals nor improves factual grounding, and increasing it can make answers less predictable.
A custom entity recognizer extracts named entities. It does not retrieve relevant passages and generate cited answers from an evolving document collection.
How to reason through AIF-C01 scenarios
AIF-C01 is a practitioner exam, not an implementation exam. AWS describes the target candidate as someone who uses AI services and understands business applications but does not necessarily build models, tune hyperparameters, or implement ML infrastructure. Start each scenario by identifying the business task, the type of data, whether the requirement is predictive or generative, and the controlling constraint—accuracy, explainability, freshness, safety, security, latency, or cost. For traditional AI services, match modality to service: text analysis points to Comprehend, speech-to-text to Transcribe, text-to-speech to Polly, translation to Translate, conversational interfaces to Lex, document extraction to Textract, and image or video analysis to Rekognition. Choose SageMaker AI when the organization needs to build or operate custom ML models. Choose Amazon Bedrock when the requirement is to build generative-AI applications with foundation models through managed APIs. For foundation-model applications, separate the levers. Prompt engineering changes the instructions supplied at inference time. Retrieval-augmented generation supplies current or private evidence. Fine-tuning changes model behavior from examples. Guardrails apply content and safety policies. Model evaluation compares quality against relevant task and business metrics. For governance questions, distinguish preventive controls from detective controls: IAM and Guardrails restrict access or content; encryption protects data; CloudTrail and model invocation logging provide audit evidence. Always eliminate options that act at the wrong layer before comparing products.
AIF-C01 field notes: if this, choose that
| Scenario clue | Choose | Avoid | Why |
|---|---|---|---|
| Predict a numeric value such as demand or price | Regression | Classification | Regression produces a continuous numeric output. |
| Assign a known label such as fraud or not fraud | Classification | Clustering | Classification predicts among predefined classes. |
| Find natural groups without labeled examples | Clustering | Supervised learning | Clustering is an unsupervised learning task. |
| Analyze sentiment, entities, or key phrases in text | Amazon Comprehend | Amazon Rekognition | Comprehend is the managed NLP service for text analysis. |
| Extract printed text, handwriting, forms, or tables from documents | Amazon Textract | Amazon Transcribe | Textract extracts structured information from document images and PDFs. |
| Convert speech to text | Amazon Transcribe | Amazon Polly | Transcribe performs automatic speech recognition; Polly synthesizes speech. |
| Detect objects, faces, labels, or unsafe content in images and video | Amazon Rekognition | Amazon Comprehend | Rekognition handles visual analysis. |
| Build with managed foundation-model APIs | Amazon Bedrock | Training a custom model first | Bedrock provides managed access to foundation models and GenAI features. |
| Build, train, tune, or deploy a custom ML model | Amazon SageMaker AI | Amazon Bedrock alone | SageMaker AI supports the custom ML lifecycle. |
| Changing private facts with citations | RAG / Bedrock Knowledge Bases | Frequent fine-tuning | Retrieval supplies current evidence without changing model weights. |
| Teach stable style, format, or task behavior from examples | Fine-tuning | RAG as the only change | Fine-tuning changes durable model behavior; retrieval supplies knowledge. |
| Block harmful topics, denied topics, or sensitive information | Bedrock Guardrails | CloudTrail | Guardrails acts on prompts and responses; CloudTrail is an audit service. |
| Who invoked an AWS API and when | AWS CloudTrail | Amazon CloudWatch metrics | CloudTrail records AWS API activity and caller identity. |
| False negatives are most costly | Recall | Accuracy alone | Recall measures how many actual positives the model finds. |
| False positives are most costly | Precision | Recall alone | Precision measures how many predicted positives are correct. |
| Classes are imbalanced and precision and recall both matter | F1 score | Accuracy alone | F1 balances precision and recall and is more informative than accuracy alone. |
| Need repeatable, consistent factual output | Lower temperature | Higher temperature | Lower temperature reduces sampling variability. |
| Need more varied brainstorming output | Higher temperature | Treating temperature as added knowledge | Temperature changes randomness, not what the model knows. |
| Restrict which principal can invoke a model | IAM least privilege | Prompt instructions | Authorization must be enforced at the access-control layer. |
| Explain which input features influenced a prediction | SageMaker Clarify | CloudWatch | Clarify supports bias analysis and feature attribution. |
Ready for all 500+ practice questions?
Sign up free to practice the complete AWS Certified AI Practitioner question bank and save your progress. Or take the shorter free practice test without signing up.
Practice 5 similar questions
Same cert, same or adjacent domain. Use these after reviewing the explanation.
Related AWS Certified AI Practitioner Practice Questions
Quick FAQ
What are the AIF-C01 domain weights?
AWS currently lists AI and ML fundamentals at 20%, generative AI fundamentals at 24%, foundation-model applications at 28%, responsible AI at 14%, and security, compliance, and governance at 14% of scored content.
How many questions are on the AWS AI Practitioner exam?
The official exam guide says the exam includes 50 scored questions and 15 unscored questions. Unscored questions are not identified. The minimum passing scaled score is 700 out of 1,000.
Does AIF-C01 require coding or model training?
The official target profile focuses on using AI and understanding practical business applications. Developing algorithms, feature engineering, hyperparameter tuning, and implementing ML infrastructure are listed as out of scope.
Are these recalled AWS exam questions?
No. CloudYeti uses original practice scenarios based on the public exam guide and AWS documentation. It does not publish dumps or recalled live exam content.
Official sources used
Verify program and exam details at the source before scheduling.