50 Oracle Cloud Infrastructure 2025 AI Foundations Associate Practice Questions: Question Bank 2025

A product team wants to compare two binary classifiers for fraud detection. They care most about avoiding missed fraud cases (false negatives), even if it increases false positives. Which metric should they prioritize?

You need an OCI managed service that extracts text and key-value information from scanned forms such as invoices and receipts. Which OCI AI Service is the best fit?

A stakeholder asks why a chatbot sometimes produces confident-sounding but incorrect answers. What is the most accurate term for this behavior in large language models?

You are designing an AI solution and must choose between supervised and unsupervised learning. You have historical data with known outcomes (labels) such as "churned" or "not churned." Which learning type is most appropriate?

A retail company wants to analyze customer feedback in multiple languages to determine whether comments are positive, negative, or neutral. Which OCI AI Service capability best addresses this requirement?

You trained a model and observe 98% accuracy, but the dataset contains 98% "non-fraud" and 2% "fraud". The model predicts everything as "non-fraud." What is the key issue and the best next diagnostic to use?

A team wants to reduce hallucinations in an LLM-powered assistant by ensuring responses are grounded in the company’s internal knowledge base stored outside the model. Which architecture pattern best fits?

Your OCI-based application processes customer images for analysis. Security requires that only a specific microservice can call the OCI Vision API, and no user should have direct permissions to the service. What is the recommended OCI approach?

You fine-tune a generative model on internal support tickets. After deployment, it sometimes outputs snippets that look like exact training examples, including sensitive data. Which mitigation is most appropriate to reduce this risk?

A data scientist builds a model using a feature derived from "last 30 days of transactions". During training, the feature was computed using the full dataset including future transactions relative to each record’s timestamp. In production, the model performs much worse. What is the most likely cause?

A retail company wants to analyze customer reviews to determine whether feedback is positive, negative, or mixed. They do not want to build or train a custom model. Which OCI AI Service is the best fit?

In machine learning, what is the primary purpose of a validation dataset?

Which statement best describes the difference between AI, machine learning (ML), and deep learning (DL)?

A security team wants to expose an internal REST endpoint that calls OCI Generative AI. They need to restrict which clients can call it, avoid embedding long-lived credentials in code, and allow permissions to be managed centrally. Which approach is most aligned with OCI best practices?

A data scientist reports that a classification model achieves 98% accuracy, but almost all predictions are the majority class and the minority class is rarely detected. Which evaluation metric is most appropriate to assess performance on the minority class?

A team is building a retrieval-augmented generation (RAG) chatbot using OCI Generative AI plus enterprise documents. They observe that answers are often off-topic because too many irrelevant passages are retrieved. Which change is most likely to improve retrieval relevance?

You deployed an OCI Vision image classification workflow, but results are inconsistent across repeated runs on the same input images. There is no model retraining between runs. Which explanation is most plausible?

A model performs extremely well on training data but poorly on new, unseen data. Which technique is most directly used to reduce this issue?

A bank plans to use an LLM to draft responses to customer emails. They must reduce the risk of data leakage and ensure sensitive identifiers (account numbers) are not returned in generated text. Which combination is the most appropriate mitigation strategy?

You are designing an ML pipeline on OCI where training data is stored in Object Storage and must be accessed by training jobs running on OCI services without distributing user credentials. Security requires least privilege and auditability. Which design best meets these requirements?

You are explaining model performance to a business stakeholder. Which statement best describes the difference between training and inference?

A team wants to detect objects in images (for example, finding cars and pedestrians) without building custom deep learning infrastructure. Which OCI AI Service is the best fit?

A data scientist says their model has high bias. Which symptom most strongly indicates high bias?

A bank wants to build a chatbot that answers questions using internal policy PDFs. They want to reduce hallucinations by grounding responses in those documents. Which approach is MOST appropriate?

A team is building a multi-step AI workflow on OCI. Step 1 transcribes customer calls, Step 2 extracts key phrases, Step 3 stores results in a database. They want reliable orchestration, retries, and visibility across steps. Which OCI service is best suited to orchestrate this workflow?

A model is trained to classify support tickets into categories. The dataset is highly imbalanced: 90% are 'General' and 10% spread across many other categories. Accuracy looks high, but users complain the model misses rare categories. Which metric is MOST appropriate to highlight this issue?

A company uses OCI Language to analyze customer feedback. Results are inconsistent across runs when processing the same text in a batch pipeline. The team suspects nondeterministic model behavior. What is the MOST likely explanation?

You are deploying a generative AI app that must prevent the model from revealing sensitive information included in prompts (for example, account numbers) and must also reduce prompt-injection risk. Which combination is the BEST approach?

A team trains an ML model and observes excellent validation performance. Later, in production, performance drops sharply. They discover that a feature was computed using information only available after the prediction time (for example, a 'resolved_date' field). What is the underlying issue?

An enterprise wants to let multiple teams use OCI generative AI capabilities while ensuring data isolation and preventing one team from accessing another team’s prompts, embeddings, or document stores. Which design is MOST aligned with OCI best practices?

A retail team is evaluating a binary classifier. On a test set, the model achieves 95% accuracy, but only 10% of the examples are positive (fraud). Which metric is generally more informative for assessing how well the model identifies the positive class?

You need to analyze customer support feedback and automatically identify entities such as product names, locations, and people, without building a custom model. Which OCI AI Service best fits this requirement?

In a generative AI workflow, which technique most directly reduces hallucinations by forcing the model to ground responses in enterprise documents at query time?

A data scientist reports that their model performs much better on the training set than on the validation set. Which issue is the most likely cause?

A bank wants to implement chat-based self-service in OCI but must prevent the model from exposing confidential internal policy text and must limit responses to only approved content. Which approach is the best practice?

You are building an invoice processing pipeline. The key requirement is to extract structured fields (invoice number, total amount, vendor name) from PDF invoices with varying layouts. Which OCI AI Service is designed for this task?

An ML engineer is preparing a dataset with 500 numerical features of very different scales (some in dollars, some in milliseconds, some in percentages) for a gradient-based algorithm. Which preprocessing step is most appropriate to improve training stability and convergence?

A team wants to evaluate an LLM-powered assistant. They notice that answers are correct but often omit citations to the source documents provided via RAG. Which change most directly increases the likelihood that the model includes citations in its responses?

A company wants to deploy an ML model in OCI and ensure that only authorized applications can invoke the model endpoint, without embedding long-lived credentials in code. Which OCI security approach best meets this requirement?

You are diagnosing poor retrieval quality in a RAG solution. Users ask highly specific questions, but the retrieved passages are frequently off-topic even though the source corpus is relevant. Which change is most likely to improve retrieval relevance?

A product team is evaluating two models. Model A has higher accuracy but is slow and expensive to run. Model B has slightly lower accuracy but meets latency requirements for an interactive application. Which metric is MOST directly used to decide whether the model meets the user experience requirement?

You want to use an OCI AI Service to extract key-value pairs and tables from uploaded invoices that may be scanned and rotated. Which OCI AI Service is the best fit?

A data scientist notices a model performs very well on the training set but significantly worse on the validation set. What is the MOST likely issue?

A retail company wants to build a conversational assistant using an LLM but must reduce the risk of the model fabricating product details. Which approach is the BEST practice to improve factual grounding?

You are designing a fraud detection model where only 0.5% of transactions are fraudulent. Accuracy is very high even for a naive model that predicts 'not fraud' for all cases. Which metric is MOST appropriate to evaluate model performance for the fraud class?

A team uses OCI Vision to classify images. They report frequent errors when photos are taken in low light and at unusual angles. Which action is MOST likely to improve model robustness?

You are integrating an OCI AI Service into an application hosted on OCI. Security requires that the application uses dynamic credentials and avoids long-lived API keys. Which authentication approach best meets this requirement?

A support team wants to automatically categorize incoming customer emails into topics (billing, technical issue, cancellation) using an OCI AI Service, but they do not want to build a custom ML pipeline. Which service capability is the MOST appropriate?

Your organization uses RAG with an LLM for internal policy Q&A. Users complain that answers are well-written but sometimes cite irrelevant policy sections. You suspect retrieval quality issues. Which change is MOST likely to improve retrieval relevance?

A model is deployed and monitored. Over several weeks, the distribution of input features (for example, customer device types and geographies) shifts significantly, and performance degrades. Which concept BEST describes what is happening, and what is the appropriate response?