CertArea
All Oracle exams

Free Oracle Cloud Infrastructure 2026 AI Foundations Associate practice test: 30 questions with answers

30 original questions for Oracle Cloud Infrastructure 2026 AI Foundations Associate, drawn from every topic area. Answer each one to see whether you got it right and why. No sign-up.

Free questions
30
Full bank
110 questions

Oracle Cloud Infrastructure 2026 AI Foundations Associate exam guide: topics, cost and format

All 30 Oracle Cloud Infrastructure 2026 AI Foundations Associate questions with answers

The same questions as the test above, grouped by exam topic. Open any question to see the answer and explanation.

Generative AI and Large Language Models (8 questions)

  1. Question 1

    What is a large language model (LLM)?

    1. AA visualisation tool that draws charts of large datasets for the business users in the reporting team each month
    2. BA database that stores very large files for the customer's use and provides fast retrieval of the records stored in them
    3. CA network device that routes very large volumes of traffic between data centres for the cloud provider's own infrastructure
    4. DA model trained on large amounts of text to predict and generate language, which can answer questions and summarise content
    Show answer and explanation
    Correct answer: D

    An LLM is trained on large text corpora to predict and generate language. Storage, network devices and charting tools are different technologies.

  2. Question 2

    What is a prompt when working with a generative AI model?

    1. AThe input text or instructions given to a generative model to guide its response
    2. BA warning message from the database that a query has taken longer than expected to complete during its run
    3. CThe output of a model after it has been fine-tuned on the customer's own dataset for a new task in the project
    4. DA password entered by the user to access the generative AI service in the cloud account on each sign-in
    Show answer and explanation
    Correct answer: A

    A prompt is the input or instructions that guide a generative model. Database warnings, fine-tuned outputs and passwords are unrelated.

  3. Question 3

    What does retrieval-augmented generation (RAG) do?

    1. AIt retrains the entire model each time a user asks a question, so that the model always has the newest knowledge available to it
    2. BIt retrieves relevant documents from a knowledge source and includes them in the prompt, so the model answers from that content
    3. CIt removes the need for any knowledge source, since the model memorises all of the documents during its training run on them
    4. DIt replaces the model's output with a fixed list of documents shown to the user in the order that they were originally stored
    Show answer and explanation
    Correct answer: B

    RAG retrieves relevant content and places it in the prompt so answers are grounded in it. Retraining per question, memorising everything and replacing output with document lists are not RAG.

  4. Question 4

    What is a hallucination in a generative AI system?

    1. AA system alert that warns the team that the training job has used more memory than was allocated for it by the platform
    2. BA visual effect in a chart that makes the data appear three-dimensional on the reporting dashboard used by the team
    3. CA confident output that is not supported by the source data or facts, and which may be false
    4. DA feature that lets the model generate images from a text description provided by the user at the start of a session
    Show answer and explanation
    Correct answer: C

    A hallucination is a confident output that is unsupported or false. Chart effects, memory alerts and image generation are unrelated.

  5. Question 5

    Which setting helps a model give more controlled or consistent answers?

    1. ARaising the temperature parameter to its highest value, so that the model varies its output as much as it possibly can
    2. BLowering the temperature parameter, so the model chooses more likely tokens and produces more consistent output
    3. CIncreasing the number of tokens in every prompt, so that the model has more to think about each time it is asked a question
    4. DTurning off the tokeniser, so that the text is not split into units for the model to process in its input layer at all
    Show answer and explanation
    Correct answer: B

    Lower temperature makes sampling more deterministic, giving more consistent output. Higher temperature increases variation, while longer prompts and disabling the tokeniser do not make answers more controlled.

  6. Question 6

    What is a token in the context of language models?

    1. AA unit of text, such as a word or a part of a word, that the model processes as input or output
    2. BA payment credential used to pay for each call to the language model in the cloud account each month of the year
    3. CA digital signature that proves the model's output was generated by the approved version of the model in the environment
    4. DA login session that expires after a period of inactivity in the generative AI console used by the staff each working day
    Show answer and explanation
    Correct answer: A

    Tokens are the units of text a language model reads and writes. Payment credentials, signatures and login sessions are different things.

  7. Question 7

    Which practice improves the quality of prompts for a language model?

    1. AWriting the prompt in capital letters only, which signals to the model that the request is important to the person asking it
    2. BStating the task clearly, giving context and examples, and specifying the format of the desired output
    3. CKeeping the prompt as short as possible, with no context at all, so that the model is not distracted by additional details
    4. DRepeating the same prompt many times in one request, which the model learns from in order to improve its eventual answer
    Show answer and explanation
    Correct answer: B

    Clear tasks, context, examples and output format guide the model to better answers. Capitals, minimal prompts and repetition do not improve the quality of the response.

  8. Question 8

    Which two statements about prompting and fine-tuning are correct? (Choose two.)

    1. APrompting changes the model's input without changing the model's weights
    2. BFine-tuning changes the model's weights using additional training examples for a specific task or domain
    3. CPrompting requires training the model on new data for each user question that is asked of the system
    4. DFine-tuning is needed for every question a user asks the model in production, since the answers are otherwise unreliable
    5. EPrompting permanently alters the model's knowledge for all future users of the system in the organisation
    Show answer and explanation
    Correct answer: A, B

    Prompting shapes the input without changing weights, and fine-tuning updates weights with examples. Retraining per question, mandatory fine-tuning and permanent prompt changes are all incorrect.

AI and Machine Learning Fundamentals (8 questions)

  1. Question 9

    Which type of machine learning learns from labelled examples in order to predict a known output?

    1. ASupervised learning, which learns a mapping from labelled inputs to known outputs
    2. BUnsupervised learning, which finds structure in unlabelled data without any known output that the model must predict
    3. CReinforcement learning, which learns from rewards and penalties received after taking actions within an environment
    4. DClustering alone, which groups similar items together but never predicts a labelled outcome from the examples it is given
    Show answer and explanation
    Correct answer: A

    Supervised learning uses labelled examples to learn how inputs map to known outputs. Unsupervised learning and clustering work without labelled outputs, and reinforcement learning learns from rewards rather than labels.

  2. Question 10

    What is overfitting?

    1. AA model that fits the training data too closely and therefore performs poorly on new, unseen data
    2. BA model that is too simple to capture the patterns present in either the training data or in new data that it sees later
    3. CA model that has been trained on too little data, so that it has not yet had the chance to learn the training patterns
    4. DA model whose predictions are always correct on the data it sees, which is taken as a sign of excellent performance in use
    Show answer and explanation
    Correct answer: A

    Overfitting means memorising the training data so that generalisation to new data suffers. Underfitting, too little training and perfect training results are different situations.

  3. Question 11

    Which metric is most suitable for evaluating a regression model that predicts house prices?

    1. AMean absolute error or root mean squared error, which measure the size of the differences between predicted and actual values
    2. BAccuracy, which measures the proportion of correct class labels and suits classification tasks rather than numeric prediction tasks
    3. CPrecision, which measures the share of positive predictions that were correct in a binary classification task over the period
    4. DThe confusion matrix count of true positives, which is designed for clustering problems rather than for numeric prediction tasks
    Show answer and explanation
    Correct answer: A

    Regression is evaluated with error measures on numeric predictions such as mean absolute error or RMSE. Accuracy, precision and true-positive counts are classification measures.

  4. Question 12

    What is the purpose of splitting data into training and test sets?

    1. ATo evaluate the model on data it did not learn from, which gives an honest estimate of its performance
    2. BTo make the training run faster by removing half of the records from the dataset before the model learns from any of them
    3. CTo store the model's predictions in a separate table for the reporting team to review each week during the reporting cycle
    4. DTo remove the labels from the data so that the model cannot see the answers during its training run on the data
    Show answer and explanation
    Correct answer: A

    A held-out test set measures performance on data the model has not learned from. Removing records to speed training, storing predictions and removing labels do not give that honest estimate.

  5. Question 13

    Which neural network type is most commonly used for image data?

    1. ALinear regression with one input variable, which is used to classify the pixels of photographs directly in the system
    2. BRecurrent networks with no memory of earlier inputs, which are used mainly for storing static tables of values
    3. CDecision trees with a single split per node, which are used to draw diagrams of images for documents in the office
    4. DConvolutional neural networks, which apply learned filters across spatial data such as images
    Show answer and explanation
    Correct answer: D

    Convolutional networks learn spatial filters that suit images. Recurrent networks, single-split trees and single-variable regression are not the usual choice for images.

  6. Question 14

    What is a feature in a machine learning dataset?

    1. AThe label or target value that the model is trained to predict for each record in the dataset as a whole
    2. BAn input variable used by the model to make a prediction
    3. CThe name of the algorithm chosen for training the model in the project each time that the training job is run
    4. DThe file format in which the training data is stored on the disk of the storage service used by the team
    Show answer and explanation
    Correct answer: B

    A feature is an input variable the model uses to predict. The label is the target, while algorithm names and file formats are other things.

  7. Question 15

    Which two statements about machine learning are correct? (Choose two.)

    1. AModel performance should be checked on held-out data before a model is used in production
    2. BA model that is accurate on its training data is always ready for production without any testing
    3. CMore relevant, good-quality training data generally helps a model generalise to new cases
    4. DDeep learning is required for every machine learning task, including simple predictions made from tabular data
    5. EBias in the training data cannot affect a model's predictions once the model has been trained on that data
    Show answer and explanation
    Correct answer: A, C

    Checking performance on held-out data before production, and using relevant good-quality data to aid generalisation, are sound practice. Training accuracy alone, mandatory deep learning and immunity to bias are incorrect.

  8. Question 16

    Which two activities are part of a typical machine learning workflow? (Choose two.)

    1. APreparing and cleaning the data before training the model on it
    2. BDeploying the model to production before any evaluation has been carried out on its results at all
    3. CEvaluating the model's performance and tuning its parameters or features based on what the results show
    4. DDeleting the training data immediately after the first training run is completed, to reduce the storage costs
    5. EWriting the model's predictions on paper before the data is loaded into any computing system at all for the project
    Show answer and explanation
    Correct answer: A, C

    Data preparation and evaluation with tuning are core steps in the workflow. Deploying before evaluation, deleting data immediately and paper predictions are not part of the normal process.

OCI AI Services (7 questions)

  1. Question 17

    Which OCI AI service extracts text from images and scanned documents?

    1. AOCI Compute, which provides virtual machines but has no built-in document text extraction capability of its own
    2. BOCI Object Storage, which keeps the files that customers upload but does not read or interpret their content in any way
    3. COCI Network Firewall, which filters network traffic but does not read the content of documents passing through it at all
    4. DOCI Vision with text recognition, or Document Understanding for structured extraction from documents
    Show answer and explanation
    Correct answer: D

    OCI Vision provides text recognition and Document Understanding extracts structured data from documents. Storage, firewalls and compute do not interpret document content.

  2. Question 18

    Which OCI AI service converts spoken audio into text?

    1. AOCI Vision, which analyses images and video frames rather than audio recordings captured from the user's device or call
    2. BOCI Language, which analyses written text for sentiment and entities but does not process audio input directly from users
    3. COCI Speech, which transcribes audio into text for applications that need speech as an input to their processing
    4. DOCI Data Science, which provides notebooks for building models but is not a speech transcription service in its own right
    Show answer and explanation
    Correct answer: C

    OCI Speech performs transcription of audio. Language analyses text, Vision analyses images, and Data Science is a model-building environment.

  3. Question 19

    Which service detects sentiment and named entities in text?

    1. AOCI Vision, which analyses the visual content of images and does not process written text for sentiment in any way
    2. BOCI Speech, which converts speech to text and cannot analyse the sentiment of written documents directly in the workflow
    3. COCI Language, which provides text analysis such as sentiment, key phrases and named entity recognition
    4. DOCI Block Volumes, which provide storage for databases and do not analyse the content of text in any form at all
    Show answer and explanation
    Correct answer: C

    OCI Language performs sentiment analysis and entity recognition on text. The other services work on speech, images or storage rather than text analysis.

  4. Question 20

    Which OCI service is designed for building, training and deploying machine learning models?

    1. AOCI Load Balancer, which distributes incoming traffic across backend servers in order to maintain the availability of a service
    2. BOCI Data Science, which provides notebooks, managed compute and model deployment for machine learning work
    3. COCI DNS, which resolves domain names to IP addresses for clients that make requests to the cloud services from their devices
    4. DOCI Email Delivery, which sends transactional email from applications to the people who use those applications each day
    Show answer and explanation
    Correct answer: B

    OCI Data Science supports the full model workflow, from notebooks to deployment. Load balancing, DNS and email delivery are important cloud services but are not model-building platforms.

  5. Question 21

    Which concept describes adapting a model trained on a large dataset to a new but related task with less data?

    1. AOverfitting, which makes a model perform well only on its training data and poorly on new data that it has not seen before
    2. BTransfer learning, which adapts a model trained on a large dataset to a new but related task
    3. CFeature deletion, which removes inputs from the dataset to make the training faster on smaller hardware in the environment
    4. DData labelling, which adds the correct answers to records so that a supervised model can learn from them during training
    Show answer and explanation
    Correct answer: B

    Transfer learning reuses a pre-trained model for a related task, which reduces the data needed. Overfitting, feature deletion and labelling describe other concepts.

  6. Question 22

    Which two statements about OCI AI services are correct? (Choose two.)

    1. APre-built AI services can be called through APIs without the customer building a model from scratch
    2. BPre-built AI services require the customer to design its own neural network architecture before any call can be made
    3. CAI services can be used together with other OCI services, such as Object Storage, in a single processing workflow
    4. DAI services are available only on-premises and cannot be used from the cloud at any point in the workflow or in practice
    5. EAI services never process personal data, so no data protection controls are needed when they are used by the business
    Show answer and explanation
    Correct answer: A, C

    Pre-built services are called through APIs and combine with other OCI services in workflows. Architecture design, on-premises-only availability and exemption from data protection are incorrect.

  7. Question 23

    Which OCI AI service identifies objects in images, such as detecting products on a shelf?

    1. AOCI Language, which analyses written text for sentiment, key phrases and entities but does not process images at all
    2. BOCI Speech, which converts spoken audio into text for applications that accept voice commands from their users each day
    3. COCI Vision, which provides image analysis such as object detection and image classification
    4. DOCI File Storage, which provides shared file systems for applications but does not analyse the images stored on them
    Show answer and explanation
    Correct answer: C

    OCI Vision provides object detection and image classification. Speech, Language and File Storage do not analyse images.

Oracle Database 23ai and Vector Search (7 questions)

  1. Question 24

    What is a vector embedding?

    1. AA security token used to sign requests to the database from the application each time that it connects to the database
    2. BA vector drawing file format used for logos and icons that the design team stores in the team's shared repository for use
    3. CA list of the database tables that are linked together by foreign keys in the schema of the application database used
    4. DA numeric representation of data such as text, which captures its meaning so that similar items sit close together
    Show answer and explanation
    Correct answer: D

    Embeddings are numeric vectors that represent meaning, so that similarity can be measured. Drawing formats, table lists and security tokens are unrelated.

  2. Question 25

    What does vector search find?

    1. AFiles whose names begin with a particular prefix, which lists the files that the user has created recently in the folder
    2. BRows that contain an exact match on a text string, which is the same thing as a standard keyword search on the table
    3. CStored items whose embeddings are most similar to the embedding of the query, ranked by a similarity measure
    4. DTables that are empty, so that the database administrator can remove them from the schema during the cleanup work
    Show answer and explanation
    Correct answer: C

    Vector search ranks stored items by the similarity of their embeddings to the query. Exact keyword matches, filename prefixes and empty tables are different searches.

  3. Question 26

    Which Oracle Database release adds a native vector data type and vector search capabilities?

    1. AOracle Database 11g, which introduced the partitioning feature for large tables that are used in the reporting workloads
    2. BOracle Database 9i, which introduced the first version of the web interface that was used for database administration
    3. COracle Database 10g, which introduced grid computing for high availability across multiple servers in the data centre
    4. DOracle Database 23ai, which adds the vector data type and the vector search features to the database
    Show answer and explanation
    Correct answer: D

    Oracle Database 23ai introduced the native vector data type and AI vector search. The earlier releases named here predate those features.

  4. Question 27

    Why is vector search used alongside an LLM in a RAG solution?

    1. ATo encrypt the prompt before it is sent to the model, which protects the question from being read by anyone in transit
    2. BTo replace the LLM entirely, since vector search can generate fluent natural language answers on its own without a model
    3. CTo find the most relevant passages from enterprise data, which are then supplied to the LLM as context for its answer
    4. DTo store the model's weights, which are too large to keep on the file system of the application server in the environment
    Show answer and explanation
    Correct answer: C

    In RAG, vector search finds relevant enterprise passages that the LLM uses as context. Vector search does not generate language, encrypt prompts or store model weights.

  5. Question 28

    Which practice keeps enterprise data secure when vector search is used with an LLM?

    1. AStoring embeddings in a public bucket so that other applications can reuse them without asking for permission from the owner
    2. BSending all enterprise documents to a public model endpoint so that the model can search them more widely on the internet
    3. CApplying the same access controls to retrieved passages as to the source data, so users see only what they are allowed to see
    4. DDisabling authentication on the database so that the vector search can run faster for the application users during the day
    Show answer and explanation
    Correct answer: C

    Retrieved passages must respect the same permissions as the source data, otherwise retrieval becomes a leak. Public endpoints, public buckets and disabled authentication each expose data.

  6. Question 29

    What does a vector database index primarily optimise?

    1. AThe number of users who can sign in to the database at the same time during the peak period of business activity each day
    2. BThe speed of writing a single row to a table of business records during the batch process that runs each night in the database
    3. CThe compression of text documents before they are stored in the file system of the database server in the data centre
    4. DThe speed of similarity search over large collections of vectors
    Show answer and explanation
    Correct answer: D

    A vector index speeds up similarity search over large vector collections. Batch writes, document compression and concurrent sign-ins are different database concerns.

  7. Question 30

    Which SQL feature lets a query compute the similarity between a stored vector and a query vector?

    1. AA vector distance function, such as cosine or Euclidean distance, used in the SELECT or ORDER BY clause of the query
    2. BA string function that trims spaces from the ends of text values stored in the business tables, for the report output
    3. CA date function that returns the current time in the database session, for the audit record that is written each time
    4. DA sequence object that generates unique numbers for new rows inserted into the table by the application each day of use
    Show answer and explanation
    Correct answer: A

    A vector distance function computes similarity between vectors in SQL. String, date and sequence functions do not compare vectors.