Azure AI Fundamentals (AI-901) Practice Test

Exam AI-901: Microsoft Azure AI Fundamentals

Azure AI Fundamentals (AI-901) overview

  • 200 practice questions
  • 31 flashcards
  • 2 study guide topics
  • 40 key terms

What you'll study

  • Identify AI concepts and capabilities42.5%
  • Implement AI solutions by using Microsoft Foundry57.5%

Microsoft Certified: Azure AI Fundamentals is an entry-level Microsoft Certification earned by passing exam AI-901. Microsoft describes the candidate as someone at the beginning of a career in AI solution development who has conceptual knowledge of AI solutions in Azure plus the foundational technical skills to work with them. Familiarity with Python coding syntax and programming techniques, Azure resources, and REST APIs, SDKs, and CLIs is expected. The exam assesses two areas: identifying AI concepts and capabilities (40โ€“45%) and implementing AI solutions by using Microsoft Foundry (55โ€“60%). The first area covers responsible AI principles, model components and configuration, model selection and deployment options, and common workloads such as generative and agentic AI, text analysis, speech, and computer vision. The second area is hands-on in emphasis: prompting, deploying and interacting with a model in the Foundry portal, building lightweight chat and agent clients with the Foundry SDK, and using Foundry Tools for speech, vision, image generation, and information extraction with Content Understanding. A score of 700 or greater is required to pass. Microsoft recommends training and hands-on experience before sitting the exam, and offers self-paced learning paths, instructor-led training, and a free Practice Assessment on AI Skills Navigator. The English version was updated on April 15, 2026, and localized versions follow about eight weeks later.

Common mistakes

Treating generative AI, machine learning, and rule-based automation as interchangeable labels, so a scenario about a chatbot that composes new text gets classified as classic classification ML.

Marketing language around AI products blurs the line, and candidates map familiar terms like 'model' or 'training' onto any AI scenario without checking whether the output is generated content or a predicted label.

What to do instead

Before answering, ask two questions: does the system produce new content (text, images, code) or assign a label/score to existing data, and does it learn patterns from examples or follow hand-written rules? Generative AI maps to content creation, ML to prediction from learned patterns, and rule-based systems to deterministic logic with no learning.

Confusing computer vision capabilities, for example describing optical character recognition as detecting objects or facial recognition as image classification, when a scenario asks which workload fits.

All vision tasks accept images as input, so the input type feels like the deciding factor and candidates stop reading for the output type, which is what actually distinguishes the workloads.

What to do instead

Classify by output: text extracted from an image is OCR, bounding boxes with labels are object detection, one label per image is classification, and identity or attribute inference from a face is facial recognition. Write the output type next to each option before choosing.

Mixing up natural language processing subtasks, such as calling sentiment analysis 'key phrase extraction' or treating translation as an entity recognition task, in scenario questions.

These workloads are often delivered by the same service and described in similar language, and candidates rely on the scenario's topic rather than the exact transformation requested.

What to do instead

Anchor on the transformation: sentiment returns a polarity, key phrase extraction returns salient terms, entity recognition returns typed entities like people or places, translation returns text in another language, and question answering returns an answer span. Match the requested output, not the domain of the text.

Assuming responsible AI principles are only about avoiding offensive output, so a scenario about a hiring model with skewed training data gets answered with a content-filtering control instead of fairness mitigation.

Safety filters are the most visible responsible AI feature, and candidates collapse all six principles into one mental bucket, losing the ability to map a specific harm to the matching principle.

What to do instead

Memorize the six principles as a checklist (fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability) and attach one concrete example to each. Then identify which harm the scenario describes and select the principle and mitigation that address that harm specifically.

Choosing a model deployment option by capacity rather than by the workload's latency and data-handling needs, for example picking a provisioned deployment for an intermittent low-volume prototype.

Candidates see throughput numbers and assume bigger is better, without weighing cost model, traffic pattern, and whether the scenario needs guaranteed capacity or pay-per-use.

What to do instead

Read the scenario for traffic pattern and budget signals: steady high volume with latency guarantees points to provisioned throughput, while spiky or experimental usage points to a standard pay-as-you-go deployment. State the deciding constraint explicitly before selecting.

Confusing the roles of a model, a deployment, and an application in Foundry, such as expecting to call a model directly without a deployment or expecting a deployment to include prompt logic.

The portal presents models and deployments on adjacent screens, and candidates who have not built an end-to-end flow assume the model endpoint and the application are the same object.

What to do instead

Draw the chain once: model catalog entry, then deployment that exposes an endpoint, then application code or an agent that sends prompts and handles responses. When a question asks what is missing from a failing call, check whether a deployment exists and whether the client targets the deployment name.

Treating prompt engineering as a substitute for grounding, so a scenario requiring answers from an internal document set is solved by rewording the system prompt instead of adding retrieved context.

Prompting is taught first and feels universally applicable, and candidates underestimate how reliably a model hallucinates on private or recent facts without supplied context.

What to do instead

Ask whether the required knowledge is in the model's training data. If the answer depends on private, current, or verifiable sources, the fix is retrieval or grounding that injects the source text, with prompt wording used only to shape format and tone.

Selecting an evaluation approach that does not match the task, for example judging a summarization solution only by exact-match scoring or skipping evaluation because the output 'looks good' in one manual test.

Candidates default to the metric they remember and do not consider that generative outputs vary, so single-sample inspection and string-matching metrics both give misleading confidence.

What to do instead

Match metric to task: classification uses accuracy, precision, recall, or F1; generation quality uses similarity or model-assisted scoring plus human review; grounded systems also need groundedness and relevance checks. Plan a small labeled test set rather than one anecdote.

Ignoring content safety and abuse controls when designing a user-facing solution, then being unable to answer which layer blocks harmful input versus harmful output.

Safety is treated as an automatic platform property, so candidates do not learn that filters can be configured and that separate controls exist for input, output, and repeated abuse.

What to do instead

Learn the layers distinctly: input filtering screens prompts, output filtering screens completions, and abuse monitoring throttles or blocks patterns over time. For each scenario, identify whether the risk enters through the user or leaves through the model, then choose the corresponding control.

Answering cost and quota questions by guessing at token economics, such as assuming output tokens are free or that a larger model always costs the same per call regardless of prompt length.

Pricing details feel like trivia, so candidates skip them, but exam scenarios embed cost constraints in the prompt and expect the candidate to reason about input versus output usage.

What to do instead

Remember that billing scales with tokens processed and that input and output are priced separately, so trimming context and limiting response length both reduce cost. When a scenario states a budget or latency limit, use those numbers to eliminate options before choosing.

Exam requirements can change. Confirm current details with the exam provider.

Azure AI Fundamentals (AI-901) practice

200 questions. Each one lists the choices, the correct answer, and an explanation. With JavaScript, this list opens one question at a time. The checker at the top of this page saves progress in this browser when it is available.

  1. Question 1Identify AI concepts and capabilities

    A company wants a system that can answer customer questions about its own product manuals without a developer writing a rule for every possible question. Which AI capability does this scenario describe?

    1. Optical character recognition
    2. Rule-based expert system
    3. Natural language processing Correct answer.
    4. Robotic process automation

    Explanation

    Natural language processing lets a system interpret and generate human language, which is what answering free-form questions about manuals requires. Because the questions are unpredictable, a hand-written rule set is impractical, so the system must learn language patterns from data.

  2. Question 2Identify AI concepts and capabilities

    Which statement best describes the relationship between artificial intelligence, machine learning, and deep learning?

    1. Machine learning and deep learning are unrelated fields within artificial intelligence
    2. Deep learning contains machine learning, which contains artificial intelligence Correct answer.
    3. Machine learning contains artificial intelligence, which contains deep learning
    4. Artificial intelligence is a subset of machine learning that uses deep learning

    Explanation

    Artificial intelligence is the broadest field, machine learning is a subset of AI that learns from data, and deep learning is a subset of machine learning that uses multi-layer neural networks. The nesting runs from broadest to narrowest in that order.

  3. Question 3Identify AI concepts and capabilities

    A hospital wants a model that flags patients likely to be readmitted, using historical records that already include readmission outcomes. Which type of machine learning is this?

    1. Reinforcement learning
    2. Generative modeling
    3. Supervised learning Correct answer.
    4. Unsupervised learning

    Explanation

    Supervised learning trains on labeled examples, where each input is paired with a known outcome. Because the historical records already contain the readmission result, the model learns a mapping from patient features to that known label.

  4. Question 4Identify AI concepts and capabilities

    What distinguishes unsupervised learning from supervised learning?

    1. It cannot be used for clustering
    2. It can only be applied to numeric data
    3. It always requires a human to label each example
    4. It trains on data without predefined labels Correct answer.

    Explanation

    Unsupervised learning finds patterns in data that has no preassigned labels, such as grouping similar customers. Supervised learning, by contrast, depends on examples paired with known target values.

  5. Question 5Identify AI concepts and capabilities

    Which task is an example of computer vision?

    1. Detecting defective parts on an assembly line from camera images Correct answer.
    2. Translating a contract from English into Japanese
    3. Summarizing a long email into three bullet points
    4. Converting a speech recording into a text transcript

    Explanation

    Computer vision extracts information from images and video, such as spotting defects in pictures of parts. The other listed tasks act on speech or text rather than visual content.

  6. Question 6Identify AI concepts and capabilities

    A team needs to convert recorded customer phone calls into searchable text. Which AI workload should they choose?

    1. Image classification
    2. Key phrase extraction
    3. Speech recognition Correct answer.
    4. Anomaly detection

    Explanation

    Speech recognition transcribes spoken audio into written text, which is exactly the step needed to make calls searchable. The other options operate on images, numeric data, or already-written text.

  7. Question 7Identify AI concepts and capabilities

    Which description best matches a generative AI model?

    1. It produces new content such as text or images in response to a prompt Correct answer.
    2. It sorts records into categories defined by historical labels
    3. It measures how similar two documents are
    4. It detects unusual values in a stream of sensor readings

    Explanation

    Generative AI models learn patterns in training data and create new content, such as text, images, or code, from a prompt. Classification, similarity scoring, and anomaly detection are different kinds of tasks.

  8. Question 8Identify AI concepts and capabilities

    Which of these is a typical use of natural language processing rather than computer vision?

    1. Detecting faces in a photo library
    2. Extracting key phrases from support tickets Correct answer.
    3. Recognizing handwritten digits on scanned forms
    4. Counting vehicles in a traffic camera feed

    Explanation

    Natural language processing works with written or spoken language, so pulling key phrases out of text fits. The other tasks analyze pixels in images, which places them in computer vision.

  9. Question 9Identify AI concepts and capabilities

    A retailer wants to group customers into segments based on purchasing behavior, but has no predefined segment labels. Which approach fits best?

    1. Classification with a supervised algorithm
    2. Reinforcement learning with a reward function
    3. Regression with a supervised algorithm
    4. Clustering with an unsupervised algorithm Correct answer.

    Explanation

    Clustering is an unsupervised technique that discovers natural groupings in unlabeled data, which matches the goal of finding segments without predefined labels. Supervised approaches would need known target labels to train on.

  10. Question 10Identify AI concepts and capabilities

    A model must predict the selling price of a house from its size, age, and location. Which machine learning task is this?

    1. Regression Correct answer.
    2. Clustering
    3. Anomaly detection
    4. Classification

    Explanation

    Regression predicts a continuous numeric outcome, and a price is a continuous value. Classification instead predicts a discrete category, while clustering and anomaly detection do not predict a known numeric target.

  11. Question 11Identify AI concepts and capabilities

    An email system must decide whether each incoming message is spam or legitimate. Which task type is this?

    1. Speech synthesis
    2. Regression
    3. Clustering
    4. Binary classification Correct answer.

    Explanation

    Binary classification assigns each item to one of two categories, which matches the spam versus legitimate decision. Regression would produce a number, and clustering would not use the known spam labels.

  12. Question 12Identify AI concepts and capabilities

    What is the primary role of training data in supervised machine learning?

    1. It stores the final predictions produced by the trained model
    2. It defines the hardware used to serve the model in production
    3. It provides labeled examples the model uses to learn a mapping to the target Correct answer.
    4. It replaces the need to evaluate the model after training

    Explanation

    Training data supplies input examples paired with correct answers so the algorithm can adjust its internal parameters. The quality and representativeness of those labeled examples strongly influence how well the resulting model generalizes.

  13. Question 13Identify AI concepts and capabilities

    A safety team wants a system that learns which actions reduce accidents by receiving feedback as it interacts with a simulated environment. Which learning paradigm is described?

    1. Supervised learning
    2. Reinforcement learning Correct answer.
    3. Transfer learning
    4. Unsupervised learning

    Explanation

    Reinforcement learning trains an agent through rewards and penalties received as it acts in an environment, which is what the simulated feedback provides. Supervised and unsupervised learning instead rely on labeled or unlabeled datasets rather than interactive feedback.

  14. Question 14Identify AI concepts and capabilities

    A company wants to build a searchable index of scanned invoices by reading the printed text in the images. Which capability is required first?

    1. Sentiment analysis
    2. Optical character recognition Correct answer.
    3. Speech synthesis
    4. Recommendation scoring

    Explanation

    Optical character recognition converts text inside images into machine-readable characters, which is a prerequisite for indexing scanned invoices. Only after extraction can text search or analysis techniques be applied.

  15. Question 15Identify AI concepts and capabilities

    Which capability allows a model to determine whether a product review expresses a positive or negative opinion?

    1. Anomaly detection
    2. Object detection
    3. Sentiment analysis Correct answer.
    4. Speech recognition

    Explanation

    Sentiment analysis is a natural language processing task that classifies the attitude expressed in text as positive, negative, or neutral. The other options analyze images, audio, or numeric patterns rather than opinions in text.

  16. Question 16Identify AI concepts and capabilities

    A translation service must convert support articles from English to Spanish while preserving meaning. Which AI workload does this represent?

    1. Reinforcement learning
    2. Computer vision
    3. Anomaly detection
    4. Natural language processing Correct answer.

    Explanation

    Machine translation is a natural language processing workload because it transforms text from one human language to another. It relies on learned language patterns rather than on image analysis or reward signals.

  17. Question 17Identify AI concepts and capabilities

    Which scenario is the clearest example of an anomaly detection workload?

    1. Predicting next quarter's revenue from historical sales
    2. Grouping news articles by topic without labels
    3. Flagging a credit card transaction that differs sharply from a cardholder's normal pattern Correct answer.
    4. Generating a product description from a short prompt

    Explanation

    Anomaly detection identifies data points that deviate from expected patterns, such as an unusual card transaction. The other scenarios describe forecasting, clustering, and content generation respectively.

  18. Question 18Identify AI concepts and capabilities

    A developer uses a pretrained language model and supplies a short instruction describing the desired output for each request, without changing the model's weights. What is this practice called?

    1. Hyperparameter tuning
    2. Model fine-tuning
    3. Feature engineering
    4. Prompt engineering Correct answer.

    Explanation

    Prompt engineering is the practice of designing instructions and context so a pretrained model produces the desired response without retraining. Fine-tuning and hyperparameter tuning instead modify the model or its training configuration.

  19. Question 19Identify AI concepts and capabilities

    A team is choosing between a classification model and a regression model for a new project. Which question best determines the correct choice?

    1. How often will the model be retrained
    2. Which programming language will the team use
    3. Is the value to predict a discrete category or a continuous number Correct answer.
    4. How many rows does the training dataset contain

    Explanation

    The nature of the target variable decides the task: discrete categories call for classification, while continuous numeric values call for regression. Dataset size, language, and retraining cadence affect implementation but not this fundamental choice.

  20. Question 20Identify AI concepts and capabilities

    A document processing pipeline must extract names, dates, and organizations from free-form contracts. Which capability is most directly applicable?

    1. Image segmentation
    2. Regression
    3. Speech synthesis
    4. Named entity recognition Correct answer.

    Explanation

    Named entity recognition is a natural language processing capability that locates and classifies elements such as people, dates, and organizations in text. It is designed for exactly this kind of information extraction from unstructured documents.

  21. Question 21Identify AI concepts and capabilities

    An organization wants an AI system to summarize long reports while citing the source passages it used. Which combination of capabilities does this require?

    1. Computer vision with object detection
    2. Reinforcement learning with clustering
    3. Natural language processing with generative AI grounded in retrieved text Correct answer.
    4. Speech recognition with anomaly detection

    Explanation

    Summarizing and citing sources requires language understanding plus a generative model that is grounded in retrieved passages so its output can be traced. Computer vision, speech, and clustering do not address summarizing text with citations.

  22. Question 22Identify AI concepts and capabilities

    A team notices that a fraud model performs well on historical records but poorly on new transactions. Which explanation is most consistent with this observation?

    1. The model learned patterns that no longer match current transaction behavior Correct answer.
    2. The model is unsupervised and therefore cannot detect fraud
    3. The model was trained with too few input features to compute a prediction
    4. The model's predictions were stored in the training data

    Explanation

    A model that excels on past records but fails on new data is likely capturing patterns that have drifted away from current behavior, a form of overfitting to stale data. Retraining on recent, representative examples is the usual remedy.

  23. Question 23Identify AI concepts and capabilities

    A hospital must decide whether to deploy a model that predicts sepsis risk. Which consideration is most important beyond raw predictive accuracy?

    1. Whether the model was trained on a graphics processing unit
    2. Whether the model's errors and decisions can be explained and audited by clinicians Correct answer.
    3. Whether the model uses the newest available algorithm family
    4. Whether the training script runs faster than competing scripts

    Explanation

    In high-stakes settings, transparency and accountability matter as much as accuracy, because clinicians must be able to understand and justify decisions. Algorithm novelty, script speed, and training hardware are secondary to safety and trust.

  24. Question 24Identify AI concepts and capabilities

    A retailer wants to recommend products by finding customers with similar purchase histories and suggesting items those similar customers bought. Which capability underlies this approach?

    1. Regression forecasting
    2. Optical character recognition
    3. Similarity-based recommendation Correct answer.
    4. Speech synthesis

    Explanation

    Recommendation systems often work by measuring similarity between users or items and surfacing what comparable users chose. This relies on comparing behavioral patterns rather than on reading images, generating speech, or forecasting a numeric value.

  25. Question 25Identify AI concepts and capabilities

    A company wants a single assistant that answers questions about scanned brochures, including text embedded in images, and replies in spoken audio. Which set of capabilities must be combined?

    1. Object detection, anomaly detection, and reinforcement learning
    2. Optical character recognition, natural language processing, and speech synthesis Correct answer.
    3. Image segmentation, classification, and forecasting
    4. Speech recognition, clustering, and regression

    Explanation

    Reading text from scanned brochures requires optical character recognition, understanding and answering the question requires natural language processing, and producing spoken replies requires speech synthesis. The other combinations omit at least one essential capability.

  26. Question 26Identify AI concepts and capabilities

    Which statement accurately describes the relationship between artificial intelligence, machine learning, and deep learning?

    1. Deep learning, machine learning, and artificial intelligence are three unrelated fields.
    2. Machine learning is a subset of deep learning, which is a subset of artificial intelligence.
    3. Artificial intelligence is a subset of machine learning, which is a subset of deep learning.
    4. Deep learning is a subset of machine learning, which is a subset of artificial intelligence. Correct answer.

    Explanation

    These fields form nested categories: AI is the broad goal of building systems that perform intelligent tasks, machine learning is one approach to that goal using data-driven models, and deep learning is a specialized form of machine learning built on multilayer neural networks.

  27. Question 27Identify AI concepts and capabilities

    A system uses rules written by experts, such as 'if temperature exceeds 80 degrees, trigger a cooling alert,' without learning from data. Which approach does this describe?

    1. Reinforcement learning
    2. Expert system Correct answer.
    3. Clustering
    4. Neural network training

    Explanation

    An expert system encodes human knowledge as explicit if-then rules and reasons over them, so behavior comes from authored logic rather than statistical learning from examples.

  28. Question 28Identify AI concepts and capabilities

    Which task belongs to computer vision rather than natural language processing?

    1. Detecting pedestrians in frames from a traffic camera Correct answer.
    2. Translating a contract into another language
    3. Extracting key phrases from meeting notes
    4. Classifying the sentiment of a support ticket

    Explanation

    Computer vision deals with interpreting image and video content, so locating objects such as pedestrians in camera frames is a vision task. The other options all operate on written language.

  29. Question 29Identify AI concepts and capabilities

    Which machine learning category learns patterns from data that has no labels attached?

    1. Unsupervised learning Correct answer.
    2. Reinforcement learning
    3. Transfer learning
    4. Supervised learning

    Explanation

    Unsupervised learning finds structure in unlabeled data, for example by grouping similar records or reducing dimensionality, without any target answers supplied during training.

  30. Question 30Identify AI concepts and capabilities

    What does the term 'inference' mean when discussing a deployed machine learning model?

    1. Adjusting model weights using a training dataset
    2. Splitting a dataset into training and test portions
    3. Removing duplicate records before training
    4. Generating a prediction from new input data Correct answer.

    Explanation

    Inference is the production phase in which a trained model applies its learned parameters to fresh inputs and returns an output such as a score, label, or generated text.

  31. Question 31Identify AI concepts and capabilities

    Which capability is the primary purpose of optical character recognition?

    1. Converting speech into written transcripts
    2. Converting one spoken language into another
    3. Converting written text into spoken audio
    4. Converting text inside images into machine-readable characters Correct answer.

    Explanation

    Optical character recognition analyzes pixels in scanned or photographed documents and outputs editable, searchable text characters that software can process.

  32. Question 32Identify AI concepts and capabilities

    Which statement best describes a large language model?

    1. It schedules jobs across servers by optimizing resource queues.
    2. It predicts likely sequences of tokens based on patterns learned from large text corpora. Correct answer.
    3. It classifies images by comparing pixel histograms against stored templates.
    4. It stores a fixed table of answers that were entered by human editors.

    Explanation

    Large language models are trained on very large text collections to estimate the probability of token sequences, which lets them complete prompts, answer questions, and generate fluent language.

  33. Question 33Identify AI concepts and capabilities

    Which scenario is an example of a regression task?

    1. Grouping news articles by topic without labels
    2. Predicting the number of minutes until a delivery arrives Correct answer.
    3. Deciding whether a photo contains a cat
    4. Assigning each support ticket to one of five priority queues

    Explanation

    Regression predicts a continuous numeric quantity, and a duration in minutes is such a quantity, so estimating arrival time is regression.

  34. Question 34Identify AI concepts and capabilities

    A call center wants to measure how frequently customers mention competitor brands during conversations. Which AI workload should be applied to the audio?

    1. Recommendation followed by ranking
    2. Anomaly detection followed by regression
    3. Image classification followed by object detection
    4. Speech recognition followed by keyword extraction Correct answer.

    Explanation

    The audio must first be transcribed into text by speech recognition, and then text analytics such as keyword or entity extraction can identify brand mentions in the transcript.

  35. Question 35Identify AI concepts and capabilities

    Which pair correctly matches an AI workload with its typical input type?

    1. Natural language processing with written text Correct answer.
    2. Computer vision with audio waveforms
    3. Anomaly detection with translated subtitles only
    4. Speech recognition with scanned photographs

    Explanation

    Natural language processing consumes text such as documents, messages, and transcripts, and produces linguistic outputs like sentiment, entities, or translations.

  36. Question 36Identify AI concepts and capabilities

    An insurance company wants to automatically read the handwritten policy number on a photographed claim form. Which capability should be used?

    1. Speaker diarization
    2. Speech synthesis
    3. Handwriting recognition Correct answer.
    4. Text-to-speech translation

    Explanation

    Handwriting recognition extends character recognition to handwritten strokes, converting the photographed script into digital text that can be validated against policy records.

  37. Question 37Identify AI concepts and capabilities

    A media company wants to index video content by searching for specific spoken words. Which combination of capabilities is required?

    1. Sentiment analysis and clustering
    2. Object detection and image captioning
    3. Face detection and OCR
    4. Speech recognition and text search Correct answer.

    Explanation

    Spoken words must be transcribed from the audio track by speech recognition, and the resulting transcript is then indexed so that text search can locate the words.

  38. Question 38Identify AI concepts and capabilities

    Which description distinguishes a discriminative task from a generative task?

    1. A discriminative task assigns labels or values to inputs, whereas a generative task creates new content. Correct answer.
    2. A discriminative task requires no data, whereas a generative task requires labeled data.
    3. A discriminative task produces new content, whereas a generative task sorts inputs into categories.
    4. A discriminative task always uses images, whereas a generative task always uses audio.

    Explanation

    Discriminative models draw boundaries that map an input to a label or numeric value, such as spam or not spam, while generative models learn data distributions well enough to produce new text, images, or audio.

  39. Question 39Identify AI concepts and capabilities

    A team wants to build a knowledge base from thousands of internal wiki pages so employees can ask questions in plain language. Which AI capability is central to this solution?

    1. Question answering over a text corpus Correct answer.
    2. Image segmentation
    3. Robotic process automation
    4. Time series forecasting

    Explanation

    Question answering retrieves or reasons over a body of documents and returns relevant answers to natural language questions, which matches the goal of querying wiki content conversationally.

  40. Question 40Identify AI concepts and capabilities

    Which scenario best illustrates reinforcement learning in practice?

    1. A robot that improves its walking gait by receiving rewards for forward progress Correct answer.
    2. A model that predicts tomorrow's temperature from past readings
    3. A model that labels scanned receipts using thousands of pre-labeled examples
    4. A model that groups website visitors by browsing similarity

    Explanation

    Reinforcement learning trains an agent through trial and error, using rewards and penalties from the environment to shape a policy that maximizes long-term reward, as in learning to walk.

  41. Question 41Identify AI concepts and capabilities

    A bank wants to flag transactions that deviate sharply from a customer's normal spending pattern without having labeled fraud examples. Which approach fits best?

    1. Supervised binary classification on labeled fraud records
    2. Unsupervised anomaly detection on transaction features Correct answer.
    3. Regression on the transaction amount
    4. Optical character recognition on statements

    Explanation

    Because no labeled fraud examples are available, the system must learn what normal behavior looks like and flag deviations, which is the defining pattern of unsupervised anomaly detection.

  42. Question 42Identify AI concepts and capabilities

    A travel company wants to identify which named cities and dates appear in free-form customer emails. Which natural language processing capability applies?

    1. Entity extraction Correct answer.
    2. Speech synthesis
    3. Image captioning
    4. Text summarization

    Explanation

    Entity extraction, also called named entity recognition, locates and classifies mentions of people, places, dates, and organizations within unstructured text such as emails.

  43. Question 43Identify AI concepts and capabilities

    Which statement about responsible AI principles is accurate?

    1. Transparency means publishing the model weights for every commercial system.
    2. Privacy means collecting all available data to improve accuracy.
    3. Accountability means people remain answerable for the behavior of AI systems they deploy. Correct answer.
    4. Fairness requires that a model never makes an incorrect prediction.

    Explanation

    Accountability holds the humans and organizations that design and operate an AI system responsible for its outcomes, including having governance and oversight in place.

  44. Question 44Identify AI concepts and capabilities

    A team must choose between training a custom model and using a pretrained foundation model for a text summarization feature. Which consideration is most directly relevant to that choice?

    1. Whether the office network uses wired or wireless connections
    2. Whether sufficient labeled data and budget exist for custom training Correct answer.
    3. Whether the team prefers a particular programming font
    4. Whether the building has a backup generator

    Explanation

    Custom training demands substantial labeled data, compute, and time, whereas a pretrained model can be adapted with far less. Comparing those resource requirements is the central decision factor.

  45. Question 45Identify AI concepts and capabilities

    A model trained to classify product photos performs poorly on photos taken with a different camera in a new warehouse. Which concept does this illustrate?

    1. A distribution shift between training and deployment data Correct answer.
    2. Model overfitting to the training labels only
    3. A syntax error in the inference code
    4. Insufficient graphics memory at inference time

    Explanation

    The deployment data differs statistically from the training data because of lighting, resolution, and camera characteristics, which is a distribution shift that degrades model accuracy.

  46. Question 46Identify AI concepts and capabilities

    An organization wants an AI assistant to answer questions using only its internal policy documents and to refuse when the answer is absent. Which design choice most directly supports this requirement?

    1. Increasing the temperature setting of the model
    2. Grounding responses in retrieved document passages Correct answer.
    3. Removing all metadata from the documents
    4. Training the model on public internet text only

    Explanation

    Retrieval-augmented grounding supplies the model with relevant passages from the approved corpus at query time, so answers can be traced to those sources and unsupported questions can be declined.

  47. Question 47Identify AI concepts and capabilities

    A computer vision model detects defects on an assembly line but raises many false alarms, causing operators to ignore alerts. Which analysis best addresses the core problem?

    1. Increase the image resolution of the cameras and assume accuracy will improve.
    2. Retrain the model using only images that contain defects.
    3. Replace the model with a regression that predicts defect size.
    4. Evaluate precision and recall and tune the decision threshold to balance false positives against missed defects. Correct answer.

    Explanation

    Frequent false alarms indicate low precision at the current threshold. Examining the precision-recall tradeoff and adjusting the threshold, possibly with better training data, directly targets the alarm behavior.

  48. Question 48Identify AI concepts and capabilities

    A team plans to combine speech, vision, and language models into one assistant. Which statement about multimodal AI is most accurate?

    1. Multimodal models eliminate the need for any training data.
    2. Multimodal models can only accept text and cannot process images.
    3. Multimodal models can process more than one data type, such as text and images, within a single system. Correct answer.
    4. Multimodal models require all inputs to be converted to audio before processing.

    Explanation

    Multimodal systems accept and relate multiple modalities, for example pairing image understanding with text generation, which enables tasks such as describing a photo or answering questions about a chart.

  49. Question 49Identify AI concepts and capabilities

    A hospital deploys a model that recommends patient follow-up intervals. Clinicians cannot explain why specific recommendations appear. Which responsible AI concern is most directly implicated?

    1. Cost of graphics processing hardware
    2. Choice of programming language for the API
    3. Network latency between data centers
    4. Interpretability and transparency of model decisions Correct answer.

    Explanation

    When stakeholders cannot understand the basis of consequential recommendations, interpretability and transparency are at stake, and techniques such as feature attribution or inherently explainable models may be needed.

  50. Question 50Identify AI concepts and capabilities

    Which type of AI workload is designed to assign a label to an input such as an email or an image?

    1. Speech synthesis
    2. Classification Correct answer.
    3. Regression
    4. Clustering

    Explanation

    Classification predicts a discrete category from a fixed set of labels, such as spam versus not spam. The model learns boundaries between predefined classes during training and outputs the most likely class for new input.

  51. Question 51Identify AI concepts and capabilities

    A retailer wants to group customers into segments based on purchasing behavior without any predefined segment names. Which machine learning approach fits best?

    1. Classification
    2. Anomaly detection
    3. Clustering Correct answer.
    4. Regression

    Explanation

    Clustering is an unsupervised technique that discovers natural groupings in data when no labels exist. Because the retailer has no predefined segment names, the algorithm must infer structure from similarity alone.

  52. Question 52Identify AI concepts and capabilities

    Which computer vision capability determines the location of individual objects within an image by drawing bounding boxes around them?

    1. Optical character recognition
    2. Image classification
    3. Image generation
    4. Object detection Correct answer.

    Explanation

    Object detection both identifies what objects are present and where they are located, returning bounding box coordinates for each detected item. Plain classification only reports which class dominates the image without spatial localization.

  53. Question 53Identify AI concepts and capabilities

    Which natural language processing task converts unstructured text into a structured set of labeled entities such as people, organizations, and locations?

    1. Sentiment analysis
    2. Text summarization
    3. Named entity recognition Correct answer.
    4. Machine translation

    Explanation

    Named entity recognition scans text and tags spans that refer to specific entity types such as persons, organizations, and places. The output is structured metadata that downstream systems can store and query.

  54. Question 54Identify AI concepts and capabilities

    Which AI workload is responsible for converting recorded audio of a person speaking into written text?

    1. Key phrase extraction
    2. Speech synthesis
    3. Speech recognition Correct answer.
    4. Sentiment analysis

    Explanation

    Speech recognition, also called speech-to-text, transcribes spoken audio into a written representation. It models acoustic signals and maps them to words in the target language.

  55. Question 55Identify AI concepts and capabilities

    Which principle of responsible AI focuses on ensuring that people understand how an AI system reaches its decisions?

    1. Fairness
    2. Privacy
    3. Transparency Correct answer.
    4. Accountability

    Explanation

    Transparency requires that the behavior and decision logic of an AI system be understandable and disclosed to stakeholders. It supports informed trust because users can see why outputs were produced.

  56. Question 56Identify AI concepts and capabilities

    A developer wants to add a pretrained language model to an application without training any model from scratch. Which Azure AI service category is intended for this purpose?

    1. Azure Storage
    2. Azure Monitor
    3. Azure AI Foundry models Correct answer.
    4. Azure Virtual Network

    Explanation

    Azure AI Foundry provides access to pretrained and foundation models that developers can invoke through APIs or integrate into applications. Infrastructure services such as monitoring, storage, and networking do not supply language models.

  57. Question 57Identify AI concepts and capabilities

    A factory wants to predict the exact number of units a machine will produce next hour based on temperature, vibration, and speed readings. Which machine learning task applies?

    1. Clustering
    2. Regression Correct answer.
    3. Classification
    4. Named entity recognition

    Explanation

    Predicting a continuous numeric quantity such as a production count is a regression task. The model learns a numeric relationship between input sensor features and the target value.

  58. Question 58Identify AI concepts and capabilities

    A support team wants to route incoming tickets to the correct department automatically based on the ticket text. Which type of model output does this require?

    1. A generated paragraph summarizing the ticket
    2. A bounding box locating text in a scan
    3. A class label chosen from a fixed set of departments Correct answer.
    4. A continuous score representing ticket urgency

    Explanation

    Routing to one of several known departments is a multiclass classification problem, so the model must output one label from a fixed set. The department list defines the classes the model learns to distinguish.

  59. Question 59Identify AI concepts and capabilities

    Which scenario is best handled by an anomaly detection model rather than a standard classification model?

    1. Sorting photographs into cats and dogs using balanced labeled data
    2. Generating marketing copy from a short prompt
    3. Identifying rare equipment failures when almost all historical records are normal Correct answer.
    4. Translating product descriptions into French

    Explanation

    Anomaly detection is designed for highly imbalanced situations where the interesting events are rare and few labeled examples exist. Standard classification assumes reasonably balanced examples of each class during training.

  60. Question 60Identify AI concepts and capabilities

    A photo app must remove distracting background behind a person and keep only the subject. Which computer vision task does this describe?

    1. Optical character recognition
    2. Image classification
    3. Image segmentation Correct answer.
    4. Object detection

    Explanation

    Image segmentation assigns a class to each pixel, producing a precise mask that separates foreground from background. Classification and detection give coarser labels or boxes that cannot isolate pixels.

  61. Question 61Identify AI concepts and capabilities

    An organization wants an AI system to summarize long meeting transcripts while preserving names and figures. Which workload category does this belong to?

    1. Computer vision
    2. Natural language processing Correct answer.
    3. Anomaly detection
    4. Speech synthesis

    Explanation

    Summarization operates on text and is a natural language processing task. The transcript must first be understood as language, then condensed while retaining key facts.

  62. Question 62Identify AI concepts and capabilities

    A company wants its virtual agent to understand that a user asking 'book me a table' and another asking 'reserve a seat for dinner' mean the same thing. Which NLP capability is most relevant?

    1. Semantic understanding of intent Correct answer.
    2. Optical character recognition
    3. Image segmentation
    4. Speech synthesis

    Explanation

    Intent recognition with semantic understanding maps different phrasings to the same underlying goal. It relies on language models that represent meaning rather than matching exact keywords.

  63. Question 63Identify AI concepts and capabilities

    Which statement best describes the role of features and labels in supervised learning?

    1. Features and labels are both unlabeled inputs used for clustering
    2. Features are input attributes and labels are the known target values used for training Correct answer.
    3. Features are the target values and labels are the input attributes
    4. Features are model weights and labels are hyperparameters

    Explanation

    In supervised learning each training example pairs input attributes, called features, with the correct answer, called the label. The model learns a mapping from features to labels so it can predict labels for new feature sets.

  64. Question 64Identify AI concepts and capabilities

    A hospital wants a model that flags suspicious regions on X-ray images and also highlights the exact pixels it used. Which capability combination is required?

    1. Image classification with segmentation Correct answer.
    2. Regression with clustering
    3. Text summarization with entity recognition
    4. Speech recognition with translation

    Explanation

    The model must both decide whether a region is suspicious and localize it at pixel level, which combines classification with segmentation. Pixel-level highlighting goes beyond bounding boxes produced by detection alone.

  65. Question 65Identify AI concepts and capabilities

    Which description best captures how a generative AI model differs from a model that only ranks or labels existing content?

    1. It can only return a fixed list of predefined categories
    2. It stores the training data and returns exact copies of it
    3. It produces new content such as text or images rather than only assigning scores to inputs Correct answer.
    4. It always requires labeled training data for every possible output

    Explanation

    Generative models learn patterns in training data and synthesize new outputs such as sentences, images, or code. Discriminative models instead output labels or scores for inputs that already exist.

  66. Question 66Identify AI concepts and capabilities

    A team wants to extract the main themes discussed in thousands of customer reviews without reading each one. Which NLP technique is most appropriate?

    1. Image classification
    2. Speech synthesis
    3. Key phrase extraction Correct answer.
    4. Object detection

    Explanation

    Key phrase extraction identifies the most salient terms and topics in a body of text, which supports theme discovery at scale. It summarizes what the text is about without requiring manual reading.

  67. Question 67Identify AI concepts and capabilities

    A bank must explain to regulators why a credit model declined a specific applicant. Which responsible AI principle is most directly engaged?

    1. Reliability of hardware
    2. Transparency and explainability Correct answer.
    3. Inclusiveness of interface design
    4. Environmental sustainability

    Explanation

    Regulatory justification requires that the factors driving an individual decision be explainable to affected parties and overseers. Transparency and explainability address exactly this need to reveal how a model reached its output.

  68. Question 68Identify AI concepts and capabilities

    A team wants to add speech-to-text transcription to an application with minimal custom model development. Which type of Azure resource should they provision?

    1. An Azure load balancer
    2. An Azure AI Services multi-service or speech resource Correct answer.
    3. An Azure container registry
    4. An Azure SQL database

    Explanation

    Azure AI Services resources expose pretrained speech-to-text capabilities through APIs and keys. Provisioning such a resource avoids building and training a custom acoustic model.

  69. Question 69Identify AI concepts and capabilities

    A team compares a model that predicts a numeric delivery time with one that predicts whether a package will arrive late. Which pair of task types do these represent?

    1. Clustering and regression
    2. Classification and regression
    3. Regression and classification Correct answer.
    4. Regression and clustering

    Explanation

    Predicting a duration is regression because the output is a continuous number. Predicting a late or on-time outcome is binary classification because the output is a category.

  70. Question 70Identify AI concepts and capabilities

    A retailer wants to forecast next quarter's revenue and also understand which input factors most influenced each forecast. Which consideration matters most when selecting the model?

    1. The model must output only fixed class labels
    2. The model must support regression and provide feature attribution Correct answer.
    3. The model must generate images from prompts
    4. The model must support clustering and hide its inputs

    Explanation

    Forecasting revenue requires regression, while explaining drivers requires interpretability such as feature attribution. Selecting a model that offers both numeric prediction and attribution satisfies the combined requirement.

  71. Question 71Identify AI concepts and capabilities

    An AI system that recommends job candidates performs well overall but selects far fewer applicants from one demographic group. Which responsible AI principle is most clearly violated?

    1. Privacy
    2. Reliability
    3. Fairness Correct answer.
    4. Transparency

    Explanation

    Fairness requires that AI systems avoid biased outcomes across groups and treat people equitably. Systematic under-selection of one demographic group is a direct fairness failure even when aggregate accuracy looks acceptable.

  72. Question 72Identify AI concepts and capabilities

    A developer must decide between a pretrained foundation model and a small custom-trained classifier for tagging short internal support notes into ten known categories. Which reasoning is most sound?

    1. The choice has no effect because all models produce identical outputs
    2. A foundation model is always preferable regardless of data and latency constraints
    3. A custom classifier is often preferable because the categories are fixed and labeled examples are available Correct answer.
    4. Neither approach can handle fixed categories with labeled examples

    Explanation

    When the label set is fixed, well defined, and supported by labeled examples, a smaller supervised classifier can be cheaper, faster, and easier to evaluate. Foundation models add flexibility but also cost and unpredictability that may be unnecessary here.

  73. Question 73Identify AI concepts and capabilities

    A team builds an assistant that reads a scanned invoice, extracts the total amount, and then answers a spoken question about that amount. Which sequence of AI capabilities does this require?

    1. Anomaly detection, regression, and translation
    2. Image segmentation, speech synthesis, and summarization
    3. Speech synthesis, image generation, and clustering
    4. Optical character recognition, information extraction, and speech recognition Correct answer.

    Explanation

    The pipeline first converts the scanned document to text with optical character recognition, then extracts the amount as structured information, and finally transcribes the spoken question with speech recognition. Each stage maps to a distinct capability that the next stage depends on.

  74. Question 74Identify AI concepts and capabilities

    Which AI workload is primarily concerned with enabling a computer to identify the objects present in a photograph?

    1. Speech synthesis
    2. Computer vision Correct answer.
    3. Knowledge mining
    4. Natural language processing

    Explanation

    Computer vision workloads process images and video to derive information such as the presence and location of objects, faces, or text. Natural language processing handles written or spoken language, speech synthesis produces audio, and knowledge mining extracts insight from large document stores.

  75. Question 75Identify AI concepts and capabilities

    A team needs a model that assigns incoming emails to one of several predefined categories such as billing, support, or sales. Which type of machine learning does this represent?

    1. Regression
    2. Reinforcement learning
    3. Clustering
    4. Classification Correct answer.

    Explanation

    Classification predicts a discrete label from a fixed set of categories, which matches assigning an email to one of several known departments. Regression predicts numeric values, clustering discovers groups without predefined labels, and reinforcement learning learns through reward signals from an environment.

  76. Question 76Identify AI concepts and capabilities

    Which capability of generative AI best explains why it can draft a product description that did not exist in its training data?

    1. It copies the closest matching sentence from its training set
    2. It retrieves the exact text from a stored product catalog
    3. It applies a fixed template chosen by a human editor
    4. It produces new content based on patterns learned from training data Correct answer.

    Explanation

    Generative models learn statistical patterns from large amounts of data and use those patterns to create original output such as text, images, or code. They do not simply look up stored documents or paste sentences, which is why the result can be fluent yet novel.

  77. Question 77Identify AI concepts and capabilities

    Which statement describes the purpose of the training phase in supervised machine learning?

    1. The model is compressed so it uses less memory
    2. The model is evaluated only on data it has never seen
    3. The model is deployed to serve predictions to end users
    4. The model learns relationships between input features and known labels Correct answer.

    Explanation

    During training, a supervised algorithm adjusts its internal parameters so that its predictions match the known labels associated with the training examples. Deployment, evaluation on unseen data, and optimization for size are separate later stages of the workflow.

  78. Question 78Identify AI concepts and capabilities

    A city wants to group its neighborhoods into segments based on similar patterns of energy consumption, without deciding the segments in advance. Which machine learning approach fits best?

    1. Binary classification
    2. Named entity recognition
    3. Clustering Correct answer.
    4. Regression

    Explanation

    Clustering is an unsupervised technique that discovers natural groupings in data when no predefined labels exist. Because the city does not specify the segments beforehand, the algorithm must form groups from similarity in the consumption patterns.

  79. Question 79Identify AI concepts and capabilities

    A developer needs to build a chatbot that answers questions about a company's internal policy documents and cites the relevant passages. Which combination of capabilities is most appropriate?

    1. Natural language processing combined with information retrieval Correct answer.
    2. Computer vision combined with speech synthesis
    3. Anomaly detection combined with regression
    4. Image classification combined with object detection

    Explanation

    Answering questions about documents requires language understanding to interpret the question and generate a response, plus retrieval to locate the relevant passages in the document store. Vision, speech, anomaly detection, and regression do not address searching and comprehending written policy text.

  80. Question 80Identify AI concepts and capabilities

    An insurance company wants a model that estimates the repair cost of a vehicle claim as a dollar amount. Which machine learning task does this describe?

    1. Clustering
    2. Regression Correct answer.
    3. Text summarization
    4. Classification

    Explanation

    Regression models predict a continuous numeric output, and a dollar amount such as a repair cost is exactly that kind of value. Classification would output a category, clustering would form unlabeled groups, and summarization would condense text.

  81. Question 81Identify AI concepts and capabilities

    Which scenario is the clearest example of a computer vision workload?

    1. Detecting fraudulent patterns in a stream of card transactions
    2. Translating a contract from French into English
    3. Generating background music for a video game
    4. Reading vehicle license plates from images captured at a toll gate Correct answer.

    Explanation

    Computer vision extracts information from visual input, so locating and reading characters in toll-gate images is a vision task. Translation is a language task, fraud detection is an analytics task on tabular or transactional data, and music generation is an audio generation task.

  82. Question 82Identify AI concepts and capabilities

    A team is choosing between a pretrained language model and training a small model from scratch for summarizing support conversations. Which consideration most strongly favors the pretrained model?

    1. It already encodes broad language understanding that would be costly to learn from limited data Correct answer.
    2. It removes the need to define what a good summary looks like
    3. It eliminates all inference costs associated with running a model
    4. It guarantees factual accuracy without any evaluation or grounding

    Explanation

    A pretrained model has already learned general language patterns from large corpora, so a team can adapt it with far less data and compute than training from scratch. It still requires evaluation, grounding, and quality criteria, and it does not remove inference costs.

  83. Question 83Identify AI concepts and capabilities

    A retailer wants to detect unusual spikes in daily sales that differ from normal seasonal patterns. Which approach is most suitable?

    1. Anomaly detection Correct answer.
    2. Binary classification with labeled fraud examples
    3. Image segmentation of store shelves
    4. Linear regression on historical revenue

    Explanation

    Anomaly detection learns what normal behavior looks like and flags observations that deviate from it, which suits detecting unexpected sales spikes without labeled examples of every unusual event. Classification needs labeled examples, regression predicts expected values rather than flagging deviations, and image segmentation is unrelated.

  84. Question 84Identify AI concepts and capabilities

    A media company wants to automatically generate captions describing the content of thousands of archived photographs. Which pairing of workloads does this require?

    1. Speech recognition to transcribe audio and regression to rank the images
    2. Knowledge mining to index the images and reinforcement learning to label them
    3. Computer vision to interpret the images and natural language processing to produce captions Correct answer.
    4. Clustering to group the images and anomaly detection to flag unusual ones

    Explanation

    Captioning requires first understanding the visual content of each photograph and then expressing that understanding as natural language. Vision supplies the interpretation of the image, and language processing generates the descriptive text.

  85. Question 85Identify AI concepts and capabilities

    An organization is evaluating a generative model for drafting legal summaries. Which evaluation practice best reflects responsible AI use?

    1. Have qualified reviewers verify outputs against source documents before use Correct answer.
    2. Use the model only for cases where no source document exists
    3. Assume outputs are correct when the model reports high confidence
    4. Disable logging so that draft content is never retained

    Explanation

    Responsible use of generative AI includes human oversight, so subject-matter experts should verify generated summaries against the authoritative source before anyone relies on them. Model-reported confidence is not a reliable substitute for review, and reducing transparency through disabled logging works against accountability.

  86. Question 86Implement AI solutions by using Microsoft Foundry

    In Microsoft Foundry, which component stores the connection details and credentials that a project uses to reach an Azure AI Search index?

    1. A connection Correct answer.
    2. A deployment
    3. An evaluation run
    4. A model card

    Explanation

    Connections in Foundry hold endpoint information and authentication details so projects and agents can call external resources such as search indexes, storage accounts, or model providers without embedding secrets in code. Deployments and evaluations are unrelated resource types.

  87. Question 87Implement AI solutions by using Microsoft Foundry

    A developer wants an agent in Foundry to answer questions from a private product manual. Which Foundry capability should be configured so the agent retrieves relevant passages before responding?

    1. A content filter severity threshold
    2. A token-per-minute rate limit
    3. A prompt flow batch run
    4. Knowledge grounding with a connected index Correct answer.

    Explanation

    Knowledge grounding attaches a searchable data source, such as an Azure AI Search index, to the agent so it retrieves relevant content and includes it in the prompt. Rate limits and content filters govern traffic and safety, not retrieval.

  88. Question 88Implement AI solutions by using Microsoft Foundry

    Which Foundry concept describes a saved, versioned configuration of a generative model that client applications can call by name?

    1. A connection
    2. A model deployment Correct answer.
    3. An index
    4. A dataset

    Explanation

    A deployment exposes a specific model version through an endpoint that applications invoke by deployment name. Datasets, connections, and indexes support evaluation, authentication, and retrieval rather than model inference.

  89. Question 89Implement AI solutions by using Microsoft Foundry

    What is the primary purpose of the system message in a Foundry model request?

    1. To list the deployment's supported regions
    2. To record billing for the request
    3. To carry the end user's latest question
    4. To set the model's behavior, role, and constraints Correct answer.

    Explanation

    The system message establishes instructions such as persona, tone, and guardrails that shape every response in the conversation. User turns carry the actual question, and billing and region data are handled by the platform rather than the prompt.

  90. Question 90Implement AI solutions by using Microsoft Foundry

    A team needs consistent, repeatable measurements of a model's answer quality against a curated question-and-answer set. Which Foundry feature should they use?

    1. A content safety filter
    2. A deployment scale setting
    3. A connection to a storage account
    4. An evaluation with a dataset and evaluator metrics Correct answer.

    Explanation

    Evaluations run a model or agent over a dataset and compute metrics such as groundedness, relevance, and similarity so quality can be compared across versions. Safety filters, scaling, and connections do not produce quality scores.

  91. Question 91Implement AI solutions by using Microsoft Foundry

    Which file format does Foundry commonly use to supply a structured set of prompt-and-response examples for an evaluation dataset?

    1. JSON Lines Correct answer.
    2. A ZIP archive of images
    3. A PDF report
    4. A YAML deployment template

    Explanation

    Evaluation datasets are typically supplied as JSON Lines files, where each line is a JSON object containing fields such as the query and expected response. Archives, reports, and templates are not row-oriented evaluation data.

  92. Question 92Implement AI solutions by using Microsoft Foundry

    Which Azure service is most commonly connected to a Foundry agent to provide vector search over private documents?

    1. Azure AI Search Correct answer.
    2. Azure Monitor
    3. Azure Key Vault
    4. Azure Load Balancer

    Explanation

    Azure AI Search supports vector and hybrid search indexes that Foundry agents query to retrieve relevant passages for grounding. Monitoring, secret storage, and traffic distribution serve different purposes.

  93. Question 93Implement AI solutions by using Microsoft Foundry

    A developer wants to call a deployed chat model from application code. What must the request include to identify which deployment to use?

    1. The model's training dataset identifier
    2. The deployment name in the request URL Correct answer.
    3. The subscription billing address
    4. The content filter severity level

    Explanation

    Foundry routes inference requests by the deployment name, which appears in the endpoint URL path. Training data identifiers, billing addresses, and filter levels do not select a deployment.

  94. Question 94Implement AI solutions by using Microsoft Foundry

    An agent answers correctly but cites sources that do not actually support its statements. Which evaluation metric most directly detects this problem?

    1. Fluency
    2. Groundedness Correct answer.
    3. Token count
    4. Latency

    Explanation

    Groundedness measures whether a response is supported by the retrieved context, so unsupported citations lower this score. Fluency judges language quality, while latency and token count describe performance rather than factual support.

  95. Question 95Implement AI solutions by using Microsoft Foundry

    A team wants the model to return responses that strictly conform to a JSON schema for downstream parsing. Which Foundry feature should they enable?

    1. A system message asking for creativity
    2. Temperature increase
    3. Top-p sampling
    4. Structured outputs Correct answer.

    Explanation

    Structured outputs constrain generation so responses match a supplied JSON schema, making parsing reliable. Sampling parameters change randomness, and creative instructions do not guarantee schema conformance.

  96. Question 96Implement AI solutions by using Microsoft Foundry

    A developer sets a high temperature on a Foundry chat deployment. What is the expected effect on responses?

    1. Stricter content filtering
    2. Faster response times
    3. Shorter output with fewer tokens
    4. More varied and less deterministic output Correct answer.

    Explanation

    Temperature controls randomness in token selection; higher values make outputs more diverse and less predictable, while lower values make them more deterministic. It does not directly shorten responses, tighten filters, or reduce latency.

  97. Question 97Implement AI solutions by using Microsoft Foundry

    A team wants to compare two prompt variants on the same set of test questions. Which Foundry workflow is most appropriate?

    1. Run an evaluation over the dataset for each variant and compare metrics Correct answer.
    2. Increase the rate limit for both deployments
    3. Add both variants to a single system message
    4. Deploy both variants and check the portal home page

    Explanation

    Running the same evaluation dataset against each variant produces comparable metrics such as relevance and similarity, enabling an evidence-based choice. Rate limits and portal views do not measure prompt quality.

  98. Question 98Implement AI solutions by using Microsoft Foundry

    Which approach best reduces hallucination when an agent answers questions about internal policies?

    1. Raise the temperature to encourage broader answers
    2. Ground responses in retrieved documents and instruct the model to say when information is missing Correct answer.
    3. Remove the system message to avoid bias
    4. Increase the maximum output tokens

    Explanation

    Grounding responses in retrieved policy text and instructing the model to acknowledge missing information limits unsupported claims. Higher temperature, removing instructions, and longer outputs tend to increase rather than reduce hallucination.

  99. Question 99Implement AI solutions by using Microsoft Foundry

    A developer needs to store an API key that a Foundry connection uses to authenticate to an external service. Which practice aligns with Foundry guidance?

    1. Reference the secret from Azure Key Vault through the connection Correct answer.
    2. Include the key in the system message
    3. Paste the key directly into every prompt
    4. Commit the key to the source repository

    Explanation

    Secrets should be held in a managed secret store such as Azure Key Vault and referenced by the connection so they are not exposed in code, prompts, or repositories. Embedding keys in prompts or source control risks leakage.

  100. Question 100Implement AI solutions by using Microsoft Foundry

    An application must stream partial model output to users as it is generated. Which Foundry request setting enables this behavior?

    1. A stricter content filter
    2. A larger maximum token limit
    3. Streaming set to true Correct answer.
    4. A higher temperature value

    Explanation

    Enabling streaming causes the service to return incremental chunks as tokens are generated, improving perceived responsiveness. Sampling and token settings change content, and filters change moderation, but none deliver partial results.

  101. Question 101Implement AI solutions by using Microsoft Foundry

    A team wants an agent to call an external ticketing system to create incidents on demand. Which Foundry capability provides this?

    1. A grounding index
    2. A model evaluation metric
    3. A tool or function the agent can invoke Correct answer.
    4. A deployment scale unit

    Explanation

    Agents can invoke tools, such as functions or API actions, to perform operations beyond text generation like creating a ticket. Indexes supply knowledge, evaluations measure quality, and scale units affect capacity.

  102. Question 102Implement AI solutions by using Microsoft Foundry

    What does the groundedness evaluator in Foundry compare when scoring a response?

    1. The response against the billing tier
    2. The response against the retrieved context provided to the model Correct answer.
    3. The response against the model's parameter count
    4. The response against the deployment's region

    Explanation

    The groundedness evaluator checks whether statements in the response are supported by the context supplied to the model. Parameter counts, regions, and billing tiers are unrelated to factual support.

  103. Question 103Implement AI solutions by using Microsoft Foundry

    A developer wants the model to reason through a multi-step math problem before answering. Which prompting technique is most appropriate?

    1. A higher top-p value
    2. A single-word answer instruction
    3. Removing all examples from the prompt
    4. Chain-of-thought prompting that requests intermediate steps Correct answer.

    Explanation

    Chain-of-thought prompting asks the model to work through intermediate reasoning steps, which improves performance on multi-step problems. Demanding one-word answers or removing guidance suppresses the reasoning that such tasks require.

  104. Question 104Implement AI solutions by using Microsoft Foundry

    A developer must choose between two models for a summarization workload with tight latency requirements. Which evidence from Foundry best supports the decision?

    1. The count of datasets in the workspace
    2. The number of connections in the project
    3. The color scheme of the portal
    4. Evaluation metrics combined with measured latency for each deployment Correct answer.

    Explanation

    Selecting a model requires balancing quality metrics from evaluations with operational measurements such as latency. Workspace counts and portal appearance provide no evidence about model suitability.

  105. Question 105Implement AI solutions by using Microsoft Foundry

    An agent sometimes ignores the instruction to answer only in English. Which change most reliably enforces the constraint?

    1. Increase the maximum token limit
    2. Reinforce the constraint in the system message and validate outputs with an evaluation Correct answer.
    3. Raise the temperature to diversify phrasing
    4. Remove the grounding index

    Explanation

    Behavioral constraints are most reliably enforced through clear system instructions plus systematic validation, such as evaluations or output checks. Sampling and token settings do not enforce language rules, and removing grounding is unrelated.

  106. Question 106Implement AI solutions by using Microsoft Foundry

    A RAG agent returns irrelevant passages because the index contains documents from unrelated departments. Which remediation most directly improves answer relevance?

    1. Disable streaming responses
    2. Apply metadata filters so retrieval targets only the relevant department Correct answer.
    3. Increase the model temperature
    4. Reduce the maximum output tokens

    Explanation

    Filtering retrieval by metadata narrows the candidate documents to the relevant department, so the model receives context that matches the question. Sampling, length, and streaming settings do not change which documents are retrieved.

  107. Question 107Implement AI solutions by using Microsoft Foundry

    A team wants to fine-tune a model on proprietary examples and then serve it in Foundry. Which sequence reflects the correct workflow?

    1. Create a connection, disable content filters, then fine-tune
    2. Deploy the base model, delete the dataset, then fine-tune
    3. Run an evaluation, then deploy without training
    4. Prepare a training dataset, run fine-tuning, deploy the tuned model, then evaluate it Correct answer.

    Explanation

    Fine-tuning requires a curated training dataset; the resulting tuned model is then deployed to an endpoint and evaluated before production use. Deleting data, altering filters, or skipping training does not produce a tuned model.

  108. Question 108Implement AI solutions by using Microsoft Foundry

    An agent must answer questions over documents that change hourly. Which design keeps answers current with the least operational overhead?

    1. Retrain the base model hourly
    2. Rebuild the deployment each hour
    3. Index documents on an automated schedule and ground responses in the refreshed index Correct answer.
    4. Embed the full document set in the system message

    Explanation

    Refreshing the search index on a schedule keeps retrieved context current without changing the model, which is far cheaper than retraining. Embedding large document sets in prompts and redeploying hourly are inefficient and error-prone.

  109. Question 109Implement AI solutions by using Microsoft Foundry

    An evaluation shows high relevance but low groundedness for an agent's answers. What is the most likely cause?

    1. The content filter is too strict
    2. The deployment region is incorrect
    3. The model answers from prior knowledge instead of the retrieved context Correct answer.
    4. The retrieved context is too short

    Explanation

    High relevance with low groundedness suggests answers address the question but are not supported by the supplied context, often because the model relies on parametric knowledge. Context length, region, and filter settings do not explain this pattern.

  110. Question 110Implement AI solutions by using Microsoft Foundry

    A developer needs deterministic outputs for a classification task that must map inputs to one of five fixed labels. Which combination is most appropriate?

    1. Streaming enabled with maximum tokens raised
    2. High top-p with no system message
    3. High temperature with free-text answers
    4. Low temperature with a schema-constrained output listing the allowed labels Correct answer.

    Explanation

    Low temperature reduces randomness, and constraining output to a schema of allowed labels ensures valid, consistent classifications. High randomness, absent instructions, and streaming do not enforce a fixed label set.

  111. Question 111Implement AI solutions by using Microsoft Foundry

    A team wants to build a chat assistant in Microsoft Foundry. Which component stores the conversation history and manages tool calls during a session?

    1. A dataset
    2. A deployment token
    3. An agent Correct answer.
    4. A connection

    Explanation

    An agent in Foundry wraps a model deployment together with instructions, tools, and conversation state, so it tracks the message history and orchestrates tool calls across a session. Connections store credentials, datasets hold evaluation or training data, and tokens authenticate requests.

  112. Question 112Implement AI solutions by using Microsoft Foundry

    A developer creates a Foundry project and wants to use a model. What must be created from a model catalog entry before the model can be called from the project?

    1. A workspace
    2. A deployment Correct answer.
    3. A schema
    4. A plugin

    Explanation

    Models in the catalog become usable only after a deployment is created, which provisions an endpoint with a chosen model version and capacity settings. Application code then calls that deployment by name rather than the catalog entry itself.

  113. Question 113Implement AI solutions by using Microsoft Foundry

    Which Foundry feature lets a developer declare that an agent has access to a function the application itself executes?

    1. Tool definitions Correct answer.
    2. Token limits
    3. Content filters
    4. Evaluation datasets

    Explanation

    Tool definitions describe callable functions to the agent, including their names, parameters, and descriptions. When the model decides to call one, the response contains the call and the application runs the function and returns the result.

  114. Question 114Implement AI solutions by using Microsoft Foundry

    A developer wants to send a system message to shape an agent's behavior. Where should that instruction be placed in the request?

    1. In the assistant role message
    2. In the user role message
    3. In the tool role message
    4. In the system role message Correct answer.

    Explanation

    The system role carries persistent instructions that guide tone, constraints, and behavior for the whole conversation. User messages carry the request, assistant messages carry prior model output, and tool messages carry function results.

  115. Question 115Implement AI solutions by using Microsoft Foundry

    A team needs its Foundry chat deployment to accept image input alongside text. Which capability must the selected model support?

    1. Multimodal input Correct answer.
    2. Fine-tuning
    3. Batch processing
    4. Structured output

    Explanation

    Handling images together with text requires a multimodal model that can process more than one input modality. Fine-tuning, batch processing, and structured output concern training, scheduling, or response formatting rather than image understanding.

  116. Question 116Implement AI solutions by using Microsoft Foundry

    Which artifact in Foundry is used to store a reusable set of instructions and configuration for an agent so it can be versioned and shared?

    1. An agent definition Correct answer.
    2. A resource group
    3. A connection string
    4. An activity log

    Explanation

    An agent definition captures the model, instructions, and tool configuration as a reusable, versionable asset within the project. Connections, resource groups, and activity logs serve authentication, resource organization, and auditing respectively.

  117. Question 117Implement AI solutions by using Microsoft Foundry

    A developer needs to know how much of the context window a prompt consumes before sending it. Which concept directly determines this?

    1. Top-p
    2. Temperature
    3. Presence penalty
    4. Token count Correct answer.

    Explanation

    Prompts and responses are measured in tokens, and the model's context window is a token budget shared by input and output. Sampling parameters such as temperature, top-p, and penalties influence output variety rather than prompt size.

  118. Question 118Implement AI solutions by using Microsoft Foundry

    A developer wants to retrieve the sources an agent used when answering with a knowledge tool. What should the application do with the agent run result?

    1. Read the citation annotations Correct answer.
    2. Disable the knowledge tool
    3. Increase the model version
    4. Delete the run history

    Explanation

    When a knowledge tool grounds a response, the run result includes citation annotations that map statements in the answer to retrieved source content. Reading those annotations lets the application display references to users.

  119. Question 119Implement AI solutions by using Microsoft Foundry

    A developer wants responses to follow a consistent persona and formatting convention across many user sessions. Which configuration is most appropriate?

    1. A shorter max token limit
    2. A new deployment per session
    3. Persistent agent instructions Correct answer.
    4. A higher top-p value

    Explanation

    Instructions are attached to the agent definition and applied to every run, which makes them the right place for persona and formatting rules. Sampling parameters and token limits affect output variety and length, not consistent behavioral conventions.

  120. Question 120Implement AI solutions by using Microsoft Foundry

    An application must let users ask follow-up questions that refer to earlier turns, such as 'what about the second one?'. Which design is required?

    1. Create a new agent for each turn
    2. Send prior conversation messages with each request Correct answer.
    3. Send only the newest user message
    4. Raise the temperature on each turn

    Explanation

    Models are stateless, so references to earlier turns resolve only if the prior messages are included in the request or maintained by the agent's thread. Sending only the latest message removes the context needed to interpret pronouns and ordinals.

  121. Question 121Implement AI solutions by using Microsoft Foundry

    A developer is building an agent that must query a relational database and summarize results. Which Foundry design keeps credentials out of prompt text?

    1. Register the database as a connection and reference it from a tool Correct answer.
    2. Include the password in each user message
    3. Store the password in the agent name
    4. Paste the connection string into the system message

    Explanation

    Connections hold endpoint and credential information in project configuration so tools can authenticate without exposing secrets in prompts. Embedding secrets in messages risks leakage through logs and conversation history.

  122. Question 122Implement AI solutions by using Microsoft Foundry

    A team wants an agent to answer questions using an Azure AI Search index and also call an internal pricing API. Which approach fits Foundry best?

    1. Attach both a knowledge tool and a function tool to one agent Correct answer.
    2. Put the index contents into the system message
    3. Create two separate agents and merge their answers manually
    4. Rely on the model's pretrained knowledge for both

    Explanation

    An agent can be configured with multiple tools, so a knowledge tool can ground answers in the index while a function tool calls the pricing API when needed. Merging separate agents or embedding index content in instructions is brittle and does not scale.

  123. Question 123Implement AI solutions by using Microsoft Foundry

    A developer wants the application to receive the agent's answer incrementally so users see text immediately. Which response handling approach should be used?

    1. Consume server-sent events from the streaming response Correct answer.
    2. Disable tool calls to shorten the run
    3. Request the maximum token limit and wait
    4. Poll the run status until completion and read the final message

    Explanation

    Streaming returns partial output as server-sent events, letting the client render text as tokens arrive. Polling waits for the entire run to finish, and changing token limits or tools does not provide incremental delivery.

  124. Question 124Implement AI solutions by using Microsoft Foundry

    An agent must complete a multi-step task that involves several tool calls before producing a final answer. What should the application check to know when the run has finished?

    1. The model version string
    2. The token count exceeding one hundred
    3. The run status reaching a terminal state Correct answer.
    4. The first tool call being issued

    Explanation

    During a run the agent may issue multiple tool calls and require the application to submit outputs before continuing. The run status indicates progress and reaches a terminal state only when the final answer is ready.

  125. Question 125Implement AI solutions by using Microsoft Foundry

    A developer wants to test an agent against a fixed set of questions and compare results before and after a prompt change. Which Foundry practice supports repeatable comparison?

    1. Change the dataset between runs
    2. Delete prior evaluation results before testing
    3. Run evaluations against the same saved dataset each time Correct answer.
    4. Ask ad hoc questions manually after each change

    Explanation

    Repeatable comparison requires holding the input constant, so the same saved dataset should be evaluated before and after each change. Ad hoc questioning, deleting results, or altering the dataset removes the baseline needed for comparison.

  126. Question 126Implement AI solutions by using Microsoft Foundry

    A developer wants to limit how much a single agent response can cost in tokens. Which setting should be configured?

    1. The retrieval chunk size
    2. The number of tools
    3. The maximum output tokens Correct answer.
    4. The system message length

    Explanation

    The maximum output token setting caps how many tokens the model may generate per response, which bounds cost and latency. System message length, tool count, and chunk size influence input or retrieval behavior rather than the output cap.

  127. Question 127Implement AI solutions by using Microsoft Foundry

    An agent returns answers in a different language than requested even though the instruction says to reply in French. Which change most directly improves instruction adherence?

    1. Increase the temperature to add variety
    2. Restate the language rule explicitly in the system instructions with an example Correct answer.
    3. Shorten the user message
    4. Add more tools to the agent

    Explanation

    Instruction adherence improves when the constraint is stated clearly and unambiguously in the system instructions, ideally with an example of the expected output. Sampling changes, extra tools, and shorter user input do not strengthen the language constraint.

  128. Question 128Implement AI solutions by using Microsoft Foundry

    A knowledge-grounded agent sometimes cites a document that does not contain the claim it supports. Which remediation most directly targets this problem?

    1. Increase the number of retrieved chunks
    2. Remove the citation requirement
    3. Require the agent to quote the supporting passage when citing Correct answer.
    4. Raise the temperature for more fluent answers

    Explanation

    Requiring a verbatim supporting quote forces the model to tie each claim to text that actually exists in the source, which reduces unsupported citations. Retrieving more chunks or raising temperature does not enforce that linkage.

  129. Question 129Implement AI solutions by using Microsoft Foundry

    A team needs an agent to summarize a large batch of support tickets overnight rather than interactively. Which Foundry capability is most appropriate?

    1. Multimodal input
    2. Interactive tool calls
    3. Batch processing of requests Correct answer.
    4. Streaming responses

    Explanation

    Batch processing submits many requests for asynchronous completion, which suits high-volume overnight workloads at lower cost than interactive calls. Streaming, interactive tools, and multimodal input target real-time or rich-input scenarios.

  130. Question 130Implement AI solutions by using Microsoft Foundry

    A developer must choose a model for an agent that will handle long policy documents in a single prompt. Which model property is the primary constraint to evaluate?

    1. The number of available tools
    2. The deployment region
    3. The temperature range
    4. The context window size Correct answer.

    Explanation

    Long documents consume many tokens, so the model's context window determines whether the full text plus instructions and answer fit in one request. Temperature range, tool count, and region do not govern how much text fits.

  131. Question 131Implement AI solutions by using Microsoft Foundry

    A developer wants an agent to return a confidence score alongside a free-text summary. Which Foundry feature best supports this requirement?

    1. Lowering the top-p value
    2. Adding a second agent
    3. Structured output with a defined schema Correct answer.
    4. Increasing the max token limit

    Explanation

    Structured output constrains the response to a schema containing both a summary field and a numeric confidence field, so the application can parse them reliably. Sampling changes, extra agents, and token limits do not define response fields.

  132. Question 132Implement AI solutions by using Microsoft Foundry

    An agent's answers are fluent and relevant but frequently contradict the retrieved policy text. Which change addresses the root cause most directly?

    1. Increase the temperature to make answers less rigid
    2. Switch to a smaller model with faster responses
    3. Reduce the number of retrieved passages
    4. Instruct the agent to answer only from retrieved content and to say when it is unsure Correct answer.

    Explanation

    Contradicting retrieved content indicates the model is relying on its own priors rather than grounding in sources. Instructions that restrict answers to retrieved content and permit an uncertainty response directly target that behavior.

  133. Question 133Implement AI solutions by using Microsoft Foundry

    A developer must ensure an agent never calls a destructive internal function during testing. Which control is most appropriate in Foundry?

    1. Reduce the maximum output tokens
    2. Omit the destructive function from the agent's tool definitions in the test environment Correct answer.
    3. Use a newer model version
    4. Set the temperature to zero

    Explanation

    A tool the agent cannot see cannot be invoked, so excluding the destructive function from the test environment's tool definitions enforces the restriction deterministically. Sampling and token settings only influence output style and length.

  134. Question 134Implement AI solutions by using Microsoft Foundry

    A team observes that an agent sometimes answers without calling the knowledge tool even when the question clearly requires indexed content. Which change is most likely to improve tool selection?

    1. Clarify in the instructions and tool description when the tool must be used Correct answer.
    2. Add several unrelated tools to the agent
    3. Raise the top-p value
    4. Increase the number of retrieved chunks per query

    Explanation

    Tool selection depends on the model understanding when a tool applies, which is driven by instructions and the tool's description. Making the trigger conditions explicit improves the likelihood of a correct call.

  135. Question 135Implement AI solutions by using Microsoft Foundry

    A developer wants to compare two models on latency and answer quality using identical inputs before choosing one for production. Which approach is most rigorous?

    1. Evaluate both deployments on the same dataset and compare metrics and latency Correct answer.
    2. Choose the model with the larger context window without testing
    3. Rely on the model catalog description alone
    4. Test each model with a few different questions and pick the faster one

    Explanation

    A fair comparison holds inputs constant, so both deployments should run the same evaluation dataset while quality metrics and measured latency are recorded. Informal questioning, capacity alone, or catalog text cannot establish relative performance.

  136. Question 136Implement AI solutions by using Microsoft Foundry

    A developer is creating a new agent in Microsoft Foundry and must select the model that will process every request. Which component determines the reasoning capability and response style of the agent?

    1. The application insights workspace linked to the resource
    2. The model deployment attached to the agent Correct answer.
    3. The network security group applied to the project
    4. The storage container that holds uploaded files

    Explanation

    An agent's reasoning and response behavior come from the language model deployment it is configured to use. The deployment bundles a specific model version and capacity, so changing it changes both capability and throughput. Other project resources such as storage and monitoring do not generate answers.

  137. Question 137Implement AI solutions by using Microsoft Foundry

    A developer needs to give an agent persistent access to a private Azure Storage account so it can read documents during conversations. Which Foundry construct should be configured on the agent?

    1. A prompt template with a longer system message
    2. A connection with the appropriate authentication credentials Correct answer.
    3. A conversation history retention setting
    4. A response format that enforces JSON output

    Explanation

    External services are attached to agents through connections that store the endpoint and credentials securely. The agent then uses the connection when a tool needs to reach that service. Prompt content and output formatting cannot substitute for authenticated access to a data source.

  138. Question 138Implement AI solutions by using Microsoft Foundry

    A developer wants to add a tool that lets an agent call an existing REST endpoint described by an OpenAPI specification. Which Foundry tool type fits this requirement with the least custom code?

    1. A static few-shot prompt block
    2. An OpenAPI specification tool Correct answer.
    3. A code interpreter tool
    4. A file search tool

    Explanation

    An OpenAPI specification tool lets Foundry read the endpoint contract and expose its operations to the agent as callable actions. The agent can then invoke the API with parameters derived from the schema. File search and code interpreter serve different purposes and do not describe external HTTP operations.

  139. Question 139Implement AI solutions by using Microsoft Foundry

    A team wants to observe how many requests an agent handled and how many tokens those requests consumed over the past day. Which capability should they use in the Foundry portal?

    1. The agent playground chat pane
    2. Monitoring metrics for the project Correct answer.
    3. The instruction editor
    4. The model catalog filter list

    Explanation

    Foundry exposes operational metrics such as request counts, latency, and token usage through monitoring views tied to the project and its resources. These metrics are aggregated over time and support investigation of trends. Interactive playground panes and catalog lists show configuration rather than historical usage.

  140. Question 140Implement AI solutions by using Microsoft Foundry

    A developer is adding a knowledge source and wants the agent to return the specific passages that support its answer. Which configuration supports this behavior?

    1. Shorten the system instruction
    2. Increase the model temperature
    3. Enable citations on the knowledge tool Correct answer.
    4. Disable conversation history

    Explanation

    Knowledge tools can be configured to include citations so the response references the retrieved passages that support each claim. This makes grounding verifiable by the end user. Sampling temperature, instruction length, and history settings do not attach source references to answers.

  141. Question 141Implement AI solutions by using Microsoft Foundry

    A developer must store the endpoint and key for an external search service once and reuse it across several agents in the project. Which approach avoids duplicating secrets in each agent's configuration?

    1. Paste the key into every agent's system instruction
    2. Encode the key in each agent's name
    3. Create a project connection and reference it from each agent Correct answer.
    4. Add the key to each agent's welcome message

    Explanation

    A project connection centralizes the endpoint and credential so multiple agents can reference the same secured configuration. Rotating the credential then requires a single update. Embedding secrets in prompts, names, or messages exposes them and forces repeated maintenance.

  142. Question 142Implement AI solutions by using Microsoft Foundry

    A developer wants to test an agent interactively in the portal, adjust its instructions, and immediately resend the same message to compare outcomes. Which Foundry surface supports this workflow?

    1. The quota page
    2. The agent playground Correct answer.
    3. The resource group overview
    4. The model catalog

    Explanation

    The agent playground provides an interactive chat surface where instructions and settings can be edited and messages resent for quick comparison. It is designed for iterative development before deployment. Catalog, quota, and resource overview pages present inventory and limits rather than conversational testing.

  143. Question 143Implement AI solutions by using Microsoft Foundry

    A developer must give an agent the ability to answer questions from a set of PDF files that were uploaded to the project. Which tool should be configured?

    1. A response format set to plain text
    2. An OpenAPI tool pointing to a public weather service
    3. Code interpreter with a plotting library
    4. File search over the uploaded documents Correct answer.

    Explanation

    File search indexes uploaded documents and retrieves relevant passages so the agent can ground answers in them. It handles document parsing and chunking as part of indexing. Code interpreter, unrelated API tools, and output formatting do not provide document retrieval.

  144. Question 144Implement AI solutions by using Microsoft Foundry

    A developer is building an agent that must run several tools in sequence and decide the next step based on each result. Which Foundry capability enables this behavior?

    1. A single large prompt containing every possible answer
    2. Agent orchestration that plans and selects tool calls iteratively Correct answer.
    3. A fixed response template with placeholders
    4. A static FAQ document returned verbatim

    Explanation

    Agent orchestration lets the model plan steps, invoke tools, observe results, and decide what to do next until the task completes. This loop is what distinguishes an agent from a single model call. Static documents and templates cannot branch on intermediate results.

  145. Question 145Implement AI solutions by using Microsoft Foundry

    A developer needs the agent to return a structured object with named fields so downstream code can parse it reliably. Which configuration should be applied?

    1. Raise the maximum token limit
    2. Attach an additional knowledge source
    3. Enable streaming of partial responses
    4. Define a structured output schema for the agent response Correct answer.

    Explanation

    A structured output schema constrains the response to a defined shape with named fields and types, which makes parsing deterministic. Token limits, streaming, and extra knowledge sources change size or content but not the response structure.

  146. Question 146Implement AI solutions by using Microsoft Foundry

    A developer wants the agent to remember details a user shared earlier in the same conversation but not across separate users. Which configuration achieves this?

    1. Disable history and rely on the model's training data
    2. Store user details in the system instruction
    3. Use one shared thread for all users
    4. Maintain a separate thread per user conversation Correct answer.

    Explanation

    A thread holds the message history for one conversation, so creating a thread per user preserves context within that conversation while isolating users from each other. A shared thread would mix histories, and instructions or training data cannot hold per-user session details.

  147. Question 147Implement AI solutions by using Microsoft Foundry

    A developer must let an agent perform arithmetic and reshape tabular data returned by a tool before summarizing it. Which tool is most appropriate?

    1. File search
    2. A structured output schema
    3. Code interpreter Correct answer.
    4. An OpenAPI tool for a weather service

    Explanation

    Code interpreter runs generated code in a sandbox, which suits calculations and data transformation tasks. It can process tool output programmatically before the model writes a summary. Retrieval tools, unrelated APIs, and schemas do not execute computation.

  148. Question 148Implement AI solutions by using Microsoft Foundry

    A developer wants to control how creative or deterministic an agent's answers are without changing the model. Which parameter should be adjusted?

    1. Context window size
    2. Deployment capacity
    3. Temperature Correct answer.
    4. Retention period

    Explanation

    Temperature controls sampling randomness, so lower values yield more consistent answers and higher values yield more varied ones. Context window, capacity, and retention affect limits and storage rather than creativity.

  149. Question 149Implement AI solutions by using Microsoft Foundry

    A developer must ensure an agent can only answer from approved internal documents and should refuse questions outside that corpus. Which combination is most appropriate?

    1. A knowledge tool restricted to the approved index plus instructions to decline unsupported questions Correct answer.
    2. Code interpreter plus a structured output schema
    3. A higher temperature plus a longer welcome message
    4. A public web API tool plus streaming responses

    Explanation

    Grounding the agent in a restricted index limits the material it can retrieve, and instructions tell it how to respond when retrieval finds nothing relevant. Together they reduce answers drawn from outside the approved corpus. The other combinations add capability without constraining the source of answers.

  150. Question 150Implement AI solutions by using Microsoft Foundry

    A developer needs to call an agent from a custom web application and authenticate using the application's own identity rather than a user sign-in. Which approach should be used?

    1. Send requests without any credential
    2. Embed a shared API key in the browser code
    3. Authenticate with a managed identity assigned to the calling application Correct answer.
    4. Disable authentication on the agent endpoint

    Explanation

    A managed identity lets the calling application obtain tokens from Microsoft Entra ID without storing secrets, and the agent endpoint can authorize that identity. Browser-embedded keys are exposed to users, and unauthenticated endpoints are not permitted for protected resources.

  151. Question 151Implement AI solutions by using Microsoft Foundry

    A developer wants to evaluate whether an agent's answers are grounded in retrieved content rather than invented. Which measurement should be used?

    1. The average response length in characters
    2. The number of tools configured on the agent
    3. The deployment's provisioned throughput
    4. A groundedness metric computed over responses and their retrieved sources Correct answer.

    Explanation

    Groundedness metrics compare claims in a response against the retrieved source content to detect unsupported statements. Response length, tool count, and throughput describe other properties and cannot reveal fabrication.

  152. Question 152Implement AI solutions by using Microsoft Foundry

    A developer needs an agent to ask the user for a missing order number before it can call an order lookup tool. Which design supports this?

    1. Increase the maximum response tokens
    2. Set the tool parameter as optional and call the tool anyway
    3. Define the tool parameter as required so the agent requests it when absent Correct answer.
    4. Remove the tool and answer from general knowledge

    Explanation

    Marking a parameter as required signals to the orchestrator that the value must be collected before invocation, so the agent asks the user for it. Optional parameters allow calls with missing data, and removing the tool eliminates the lookup entirely.

  153. Question 153Implement AI solutions by using Microsoft Foundry

    A team wants the same agent configuration promoted from a development project to a production project with minimal manual rework. Which practice supports this?

    1. Copy chat transcripts between projects
    2. Recreate the agent by hand in each environment
    3. Export only the agent's welcome message
    4. Define the agent declaratively and deploy the definition through a repeatable pipeline Correct answer.

    Explanation

    Declarative agent definitions can be versioned and deployed through automation, so the same configuration is reproduced consistently across environments. Manual recreation invites drift, and transcripts or messages are not configuration artifacts.

  154. Question 154Implement AI solutions by using Microsoft Foundry

    A developer wants to limit which external domains an agent's API tool may contact during a pilot. Which control is most appropriate?

    1. Increase the model's temperature setting
    2. Add more examples to the system instruction
    3. Enable streaming responses
    4. Restrict the tool's allowed endpoints through network and connection configuration Correct answer.

    Explanation

    Outbound access is governed by network rules and by which connections and endpoints are configured for the tool. These controls constrain where requests can go. Sampling settings, prompt examples, and streaming affect output behavior rather than network reach.

  155. Question 155Implement AI solutions by using Microsoft Foundry

    A developer observes that an agent ignores a newly added tool and keeps answering from its own knowledge. Which change is most likely to correct the behavior?

    1. Remove the tool from the agent
    2. Shorten the conversation history
    3. Lower the maximum token limit
    4. Update the instructions to require the tool for questions it can answer Correct answer.

    Explanation

    Agents select tools based on their instructions and tool descriptions, so explicitly directing the agent to use the tool for relevant questions improves selection. Removing the tool eliminates the capability, and token or history settings do not influence tool choice.

  156. Question 156Implement AI solutions by using Microsoft Foundry

    A developer must reduce the risk that an agent's tool call deletes production records when a user phrases a request ambiguously. Which design is most appropriate?

    1. Allow the destructive call and log the outcome afterward
    2. Return the tool result in a structured schema
    3. Increase the deployment capacity for faster responses
    4. Require explicit confirmation before executing destructive tool operations Correct answer.

    Explanation

    Human confirmation before a destructive action prevents irreversible changes triggered by ambiguous input. Logging after execution records damage rather than preventing it, and capacity or schema settings do not gate the operation.

  157. Question 157Implement AI solutions by using Microsoft Foundry

    An agent's answers are correct but arrive slowly because every request retrieves a large number of document chunks. Which adjustment best improves latency while preserving grounding quality?

    1. Add a second knowledge source
    2. Raise the temperature to speed up sampling
    3. Tune the retrieval top-k and chunk size for the knowledge source Correct answer.
    4. Increase the maximum response tokens

    Explanation

    Retrieval settings such as the number of chunks returned and chunk size directly affect how much text is processed per request. Tuning them reduces prompt size and latency while keeping relevant evidence. Temperature, extra sources, and token limits do not target retrieval volume.

  158. Question 158Implement AI solutions by using Microsoft Foundry

    A team must prove that an agent's answers comply with a policy requiring every factual claim to reference an approved document. Which combination best satisfies the requirement?

    1. A restricted knowledge index, citations enabled, and a groundedness evaluation gate Correct answer.
    2. Streaming responses and a higher temperature
    3. A larger model deployment and a longer welcome message
    4. Code interpreter and a structured output schema

    Explanation

    Restricting retrieval to approved documents limits the evidence available, citations expose the supporting passage, and a groundedness gate blocks responses that lack support. The other combinations improve performance or formatting without verifying that claims trace to approved sources.

  159. Question 159Implement AI solutions by using Microsoft Foundry

    A developer needs to compare two prompt variants across a fixed test set and decide which yields fewer unsupported claims. Which approach is most rigorous?

    1. Ask colleagues which variant reads better
    2. Test each variant once with a single ad hoc question
    3. Choose the variant that produces shorter answers
    4. Run both variants against the same dataset and compare groundedness scores Correct answer.

    Explanation

    Holding the dataset constant and comparing a groundedness metric isolates the effect of the prompt change and gives a defensible basis for selection. Single questions, subjective impressions, and answer length are not reliable measures of factual support.

  160. Question 160Implement AI solutions by using Microsoft Foundry

    A developer must let an agent respond to events from an external system without a user initiating each conversation. Which Foundry capability should be used?

    1. Add the event text to the system instruction
    2. Rely on users to paste event details into the playground
    3. Schedule a nightly export of chat transcripts
    4. Invoke the agent programmatically through its endpoint from the event handler Correct answer.

    Explanation

    Agents expose an endpoint that applications can call programmatically, so an event handler can start a run when an external event occurs. Playground use and transcript exports require human action or happen after the fact, and instructions cannot receive live events.

  161. Question 161Implement AI solutions by using Microsoft Foundry

    A developer is creating a new generative AI application and wants to use a Microsoft Foundry model that accepts both text and images as input. Which model capability should the developer select?

    1. A multimodal model Correct answer.
    2. A classification model
    3. A text-embedding model
    4. A speech-to-text model

    Explanation

    Multimodal models in Microsoft Foundry can process more than one type of input, such as text and images, in the same request. Embedding models convert content into vectors, speech models transcribe audio, and classification models assign labels, so none of them accept mixed image and text prompts as their primary purpose.

  162. Question 162Implement AI solutions by using Microsoft Foundry

    A developer wants to send a single request to a Microsoft Foundry model deployment and receive a response. Which information is required to construct that request?

    1. The model training dataset and the compute quota
    2. The Azure region and the subscription ID only
    3. The container registry and the image tag
    4. The deployment name and the project endpoint Correct answer.

    Explanation

    Calls to a model deployment are addressed by the deployment name and authorized through the project endpoint. Training data, compute quota, region alone, and container details are not part of the inference request path.

  163. Question 163Implement AI solutions by using Microsoft Foundry

    A team wants to use a Microsoft Foundry agent to answer questions about a product manual that changes frequently. Which approach keeps answers current with the least manual effort?

    1. Index the manual in a connected knowledge source and let the agent retrieve from it Correct answer.
    2. Ask users to include the relevant page in every prompt
    3. Fine-tune the base model each time the manual changes
    4. Paste the entire manual into the agent instructions

    Explanation

    Retrieval from a connected knowledge source lets the agent ground answers in the latest indexed content without changing the model. Fine-tuning is slow and costly for frequent updates, instructions have size limits, and relying on users to supply content is unreliable.

  164. Question 164Implement AI solutions by using Microsoft Foundry

    A developer needs to give an agent a tool that calls an existing REST API. Which Microsoft Foundry feature should the developer configure?

    1. An OpenAPI specification for the API Correct answer.
    2. A prompt flow evaluation run
    3. A system-assigned managed identity only
    4. A content filter policy

    Explanation

    An OpenAPI specification describes the API operations, parameters, and schemas so the agent can call the API as a tool. Managed identity handles authentication but does not describe the API, evaluations measure quality, and content filters moderate outputs.

  165. Question 165Implement AI solutions by using Microsoft Foundry

    A developer wants to compare the outputs of two model deployments on the same set of test prompts. Which Microsoft Foundry capability is designed for this task?

    1. A deployment scale-out rule
    2. An evaluation run that scores each deployment against the test set Correct answer.
    3. A private endpoint connection
    4. A content safety filter

    Explanation

    Evaluation runs apply metrics and test data to model outputs so results can be compared objectively. Scale-out rules change capacity, content filters moderate text, and private endpoints affect network access rather than output quality.

  166. Question 166Implement AI solutions by using Microsoft Foundry

    A developer wants to control cost by limiting the number of tokens a model generates in each response. Which parameter should the developer set?

    1. Presence penalty
    2. Max output tokens Correct answer.
    3. Top-p
    4. Temperature

    Explanation

    The max output tokens parameter caps how many tokens the model may produce, which directly bounds generation cost. Temperature and top-p affect randomness, and presence penalty discourages repetition, but none of them limits response length.

  167. Question 167Implement AI solutions by using Microsoft Foundry

    A developer is choosing a model for a task that requires complex multi-step reasoning. Which model family characteristic is most relevant to this requirement?

    1. An embedding model with higher dimensions
    2. A smaller model with lower latency
    3. A speech model with more languages
    4. A larger model with stronger reasoning capabilities Correct answer.

    Explanation

    Complex reasoning generally benefits from larger, more capable models that can follow multi-step instructions. Lower latency, embedding dimensions, and language coverage address different concerns and do not primarily improve reasoning quality.

  168. Question 168Implement AI solutions by using Microsoft Foundry

    A developer wants to deploy a model in Microsoft Foundry so that it can handle sudden spikes in request volume without manual intervention. Which configuration should the developer choose?

    1. A batch deployment with a long timeout
    2. A development-only deployment
    3. Provisioned deployment with fixed capacity
    4. Standard deployment with dynamic quota Correct answer.

    Explanation

    A standard deployment with dynamic quota can scale with demand and is billed per token, which suits unpredictable spikes. Provisioned capacity is fixed and better for steady load, batch deployments process large jobs asynchronously, and development deployments are not intended for production traffic.

  169. Question 169Implement AI solutions by using Microsoft Foundry

    A developer needs to give an agent access to a model deployment without storing an API key in code. Which authentication approach should the developer use?

    1. Microsoft Entra ID with a managed identity Correct answer.
    2. An anonymous access setting on the endpoint
    3. A connection string embedded in the source file
    4. A shared key stored in a public repository

    Explanation

    Microsoft Entra ID with a managed identity lets the application obtain tokens without embedding secrets. Embedded connection strings, anonymous access, and public shared keys all expose credentials or remove protection and are not recommended.

  170. Question 170Implement AI solutions by using Microsoft Foundry

    A developer is building an agent that must call a function in the application's own code to look up customer records. Which Microsoft Foundry agent tool type should the developer use?

    1. A web search tool
    2. A code interpreter tool
    3. A function tool Correct answer.
    4. A file search tool

    Explanation

    A function tool lets the agent invoke application-defined functions and return results to the model. File search retrieves from indexed documents, code interpreter runs generated code, and web search queries the internet, so none of them calls the application's own customer lookup logic.

  171. Question 171Implement AI solutions by using Microsoft Foundry

    A developer wants to improve an agent's answers by providing a few examples of the desired input and output format in the agent instructions. Which technique is this?

    1. Retrieval augmentation
    2. Zero-shot prompting
    3. Fine-tuning
    4. Few-shot prompting Correct answer.

    Explanation

    Few-shot prompting includes example interactions in the prompt to guide the model's format and behavior. Zero-shot provides no examples, fine-tuning changes model weights through training, and retrieval augmentation supplies external content rather than format examples.

  172. Question 172Implement AI solutions by using Microsoft Foundry

    A developer notices that an agent sometimes produces different answers to the same factual question. Which generation parameter should the developer lower to make responses more deterministic?

    1. Stop sequence
    2. Temperature Correct answer.
    3. Max output tokens
    4. Frequency penalty

    Explanation

    Temperature controls randomness in token selection, so lowering it makes outputs more consistent. Max output tokens limits length, frequency penalty discourages repetition, and stop sequences end generation at a marker, but none of these directly reduces variability in factual answers.

  173. Question 173Implement AI solutions by using Microsoft Foundry

    A developer wants to add a safety layer that blocks harmful content in both user prompts and model responses. Which Microsoft Foundry feature should the developer configure?

    1. Evaluation metrics
    2. Deployment quotas
    3. Content filters Correct answer.
    4. Connection secrets

    Explanation

    Content filters inspect prompts and completions for harmful categories and can block or annotate them. Evaluation metrics measure quality, quotas govern capacity, and connection secrets store credentials, so none of them provide content moderation.

  174. Question 174Implement AI solutions by using Microsoft Foundry

    A developer needs to upload documents and create an index so that an agent can retrieve relevant passages. Which Microsoft Foundry component should the developer use?

    1. A model deployment with higher quota
    2. A managed network with a private endpoint
    3. A prompt flow with a conditional node
    4. A knowledge source connected to the agent Correct answer.

    Explanation

    A knowledge source indexes uploaded or connected content and makes it retrievable by the agent. Higher quota changes capacity, prompt flow nodes orchestrate logic, and private endpoints secure networking, but none of them create a searchable index.

  175. Question 175Implement AI solutions by using Microsoft Foundry

    A developer wants to track token usage and latency for each request to an agent. Which Microsoft Foundry capability should the developer use?

    1. A content filter policy
    2. Monitoring and metrics in the project Correct answer.
    3. A deployment scale rule
    4. A model fine-tuning job

    Explanation

    Monitoring and metrics collect operational telemetry such as token consumption and response latency. Content filters moderate text, scale rules adjust capacity, and fine-tuning changes model behavior, so none of them provide per-request usage tracking.

  176. Question 176Implement AI solutions by using Microsoft Foundry

    A developer wants to use an agent to summarize a long support transcript and extract action items as a JSON object with a fixed schema. Which feature should the developer configure?

    1. A higher temperature setting
    2. A larger max output token limit
    3. Structured output defined by a schema Correct answer.
    4. A web search tool

    Explanation

    Structured output constrains the model to return fields that match a supplied schema, which makes parsing reliable. Temperature, token limits, and web search affect creativity, length, and external grounding, but they do not enforce a JSON schema.

  177. Question 177Implement AI solutions by using Microsoft Foundry

    A developer is building an agent that must answer questions about a specific set of internal documents and should not use general world knowledge. Which instruction strategy is most appropriate?

    1. Instruct the agent to ignore the retrieved context
    2. Instruct the agent to always include a web search result
    3. Instruct the agent to answer as creatively as possible
    4. Instruct the agent to answer only from retrieved context and to say when the answer is not found Correct answer.

    Explanation

    Constraining the agent to retrieved context and requiring an explicit fallback reduces unsupported answers. Creativity instructions, mandatory web results, and ignoring context all increase the chance of answers that are not grounded in the approved documents.

  178. Question 178Implement AI solutions by using Microsoft Foundry

    A developer needs to test an agent with a fixed set of prompts and compare its answers against expected reference answers. Which Microsoft Foundry evaluation approach should the developer use?

    1. A content filter preview
    2. A single interactive chat session
    3. A deployment scale-out test
    4. A batch evaluation run with a labeled dataset Correct answer.

    Explanation

    A batch evaluation run applies metrics to many prompts and compares outputs with reference answers in a dataset. A single chat session is not repeatable at scale, scale-out tests measure capacity, and content filter previews check moderation rather than answer quality.

  179. Question 179Implement AI solutions by using Microsoft Foundry

    A developer wants to give an agent the ability to generate and run Python code to analyze a dataset during a conversation. Which tool should the developer enable?

    1. The function tool
    2. The code interpreter tool Correct answer.
    3. The file search tool
    4. The web search tool

    Explanation

    The code interpreter tool runs generated code in a sandbox so the agent can perform calculations and data analysis. File search retrieves documents, function tools call application functions, and web search queries the internet, so none of them execute arbitrary analysis code.

  180. Question 180Implement AI solutions by using Microsoft Foundry

    A developer needs to ensure that an agent's responses do not contain personal data from retrieved documents. Which approach should the developer take?

    1. Increase the temperature to vary the wording
    2. Apply a content safety policy that detects and redacts sensitive information Correct answer.
    3. Increase the max output tokens
    4. Disable the knowledge source

    Explanation

    A content safety policy can detect and redact sensitive information before it reaches the user. Temperature and token limits affect style and length, and disabling the knowledge source removes the grounding the agent needs.

  181. Question 181Implement AI solutions by using Microsoft Foundry

    A developer observes that an agent's answers are grounded in retrieved documents but the retrieval step returns many irrelevant chunks. Which change most directly improves retrieval precision?

    1. Increase the max output tokens
    2. Tune the chunk size and overlap, and add a relevance threshold Correct answer.
    3. Add more few-shot examples to the instructions
    4. Increase the model temperature

    Explanation

    Retrieval precision depends on how content is chunked and filtered, so adjusting chunk size, overlap, and a relevance threshold removes noise. Temperature, token limits, and few-shot examples affect generation rather than which chunks are retrieved.

  182. Question 182Implement AI solutions by using Microsoft Foundry

    A developer must ensure that an agent uses a specific tool only when the user explicitly requests a refund, and never otherwise. Which design approach is most reliable?

    1. Remove all other tools from the agent
    2. Raise the temperature so the agent varies its tool choice
    3. Rely on the model to infer intent from the conversation history alone
    4. Describe the tool's purpose and usage conditions precisely in the tool definition and instructions Correct answer.

    Explanation

    Precise tool descriptions and instructions tell the model when a tool is appropriate and reduce unintended calls. Higher temperature increases variability, removing other tools does not prevent misuse of the refund tool, and relying on inference alone is less reliable than explicit conditions.

  183. Question 183Implement AI solutions by using Microsoft Foundry

    A developer needs to compare two chunking strategies for a retrieval agent and determine which produces more grounded answers. Which method provides the most reliable evidence?

    1. Run the same evaluation dataset against both strategies and compare groundedness metrics Correct answer.
    2. Compare the number of tokens in each chunk
    3. Review a single answer from each strategy manually
    4. Ask the agent which strategy it prefers

    Explanation

    Applying the same evaluation dataset and groundedness metrics to both strategies produces comparable, repeatable evidence. Asking the agent, comparing token counts, or reviewing one answer each does not measure grounding quality across representative inputs.

  184. Question 184Implement AI solutions by using Microsoft Foundry

    A developer wants to reduce the risk that an agent follows instructions embedded in retrieved documents that attempt to change its behavior. Which mitigation is most appropriate?

    1. Increase the temperature to diversify responses
    2. Remove content filters to allow more context
    3. Increase the number of retrieved chunks per query
    4. Instruct the agent to treat retrieved content as data and to ignore instructions found inside it Correct answer.

    Explanation

    Prompt injection through retrieved content is reduced by instructing the model to treat documents as untrusted data rather than commands. Retrieving more chunks, increasing temperature, or removing content filters increases exposure or variability without addressing the injection risk.

  185. Question 185Implement AI solutions by using Microsoft Foundry

    A developer needs to select a model for a high-volume classification task where cost per request must be minimized and simple accuracy is sufficient. Which choice is most appropriate?

    1. A speech-to-text model
    2. A large reasoning model
    3. A multimodal model with image support
    4. A smaller, lower-cost model Correct answer.

    Explanation

    A smaller, lower-cost model is appropriate when the task is simple and cost per request is the priority. Large reasoning models, multimodal models, and speech models add capability and cost that the task does not require.

  186. Question 186Implement AI solutions by using Microsoft Foundry

    A developer wants to store the connection details and authentication settings for a model endpoint once, then reference them from several agent definitions without repeating the values. Which Microsoft Foundry concept should the developer use?

    1. A content filter policy
    2. A project connection Correct answer.
    3. A deployment token
    4. A prompt flow run

    Explanation

    A connection in a Foundry project stores an endpoint URL and its credentials so that multiple assets, such as agents and flows, can reuse the same configuration without duplicating secrets. Connections centralize credential management and make rotation easier. Tokens, runs, and filter policies serve unrelated purposes.

  187. Question 187Implement AI solutions by using Microsoft Foundry

    A developer wants an agent to answer questions over a product catalog that changes daily, and the answers must reflect the newest prices. Which approach best satisfies the requirement?

    1. Increase the model temperature so answers vary
    2. Fine-tune the model each night on the catalog
    3. Store the catalog in a search index and retrieve passages at query time Correct answer.
    4. Add the catalog to the agent's system message once at creation

    Explanation

    Retrieval at query time grounds answers in the current index contents, so daily catalog updates are reflected without retraining. Fine-tuning bakes knowledge into weights and would need repeated costly jobs, while a static system message becomes stale immediately. Temperature affects variability, not freshness.

  188. Question 188Implement AI solutions by using Microsoft Foundry

    A developer is configuring an agent in Microsoft Foundry and wants the agent to respond in the same language as the user without adding instructions to every request. Where should this behavior be specified?

    1. In the agent's instructions Correct answer.
    2. In the project's connection settings
    3. In the model deployment's rate limit
    4. In the search index schema

    Explanation

    Agent instructions define persistent behavior, persona, and constraints that apply to every conversation turn, so language mirroring belongs there. Rate limits, index schemas, and connection settings control capacity, data shape, and credentials respectively, none of which influence response style.

  189. Question 189Implement AI solutions by using Microsoft Foundry

    A team wants to evaluate an agent's answer quality against a set of human-written reference answers. Which evaluation approach should they use?

    1. Groundedness scoring against retrieved context
    2. Similarity scoring between generated and reference answers Correct answer.
    3. Latency percentile analysis
    4. Token usage analysis

    Explanation

    When reference answers exist, similarity metrics compare the generated response to those references and quantify how closely the agent matches expected output. Groundedness checks whether claims are supported by retrieved context rather than by a reference. Token and latency analysis measure cost and speed, not answer quality.

  190. Question 190Implement AI solutions by using Microsoft Foundry

    An agent must query an internal inventory service. The service exposes an OpenAPI specification, and the developer wants the agent to invoke it automatically when a user asks about stock levels. Which Foundry agent capability should be configured?

    1. A code interpreter tool
    2. A file search tool
    3. An OpenAPI tool Correct answer.
    4. A grounding filter

    Explanation

    An OpenAPI tool lets an agent call external HTTP services described by an OpenAPI specification, with the model selecting operations based on the user request. File search works over uploaded documents, code interpreter executes code, and grounding filters screen content.

  191. Question 191Implement AI solutions by using Microsoft Foundry

    A developer wants to confirm that an agent's answer about a warranty policy is supported by the documents that were retrieved, rather than invented. Which evaluation metric directly measures this?

    1. Groundedness Correct answer.
    2. Fluency
    3. Coherence
    4. Relevance

    Explanation

    Groundedness evaluates whether the statements in a response are supported by the provided source context, which is exactly the check for invented claims. Relevance compares the answer to the question, while fluency and coherence judge language quality and logical flow.

  192. Question 192Implement AI solutions by using Microsoft Foundry

    A developer needs to run a repeatable evaluation of an agent each time the instructions change, using the same dataset and metrics, and compare results across runs. Which Foundry capability supports this?

    1. A saved evaluation with a fixed dataset and metrics Correct answer.
    2. Manual chat testing in the playground
    3. Application Insights sampling
    4. A content filter with severity thresholds

    Explanation

    A saved evaluation pairs a dataset with defined metrics so each run is comparable over time, which is required for regression checking after instruction changes. Playground chat is ad hoc, telemetry sampling observes production traffic, and content filters block harmful content rather than score quality.

  193. Question 193Implement AI solutions by using Microsoft Foundry

    A developer wants an agent to answer questions using documents stored in an Azure AI Search index that already exists in the same subscription. Which configuration should the developer add to the agent?

    1. A custom content filter
    2. A higher tokens-per-minute quota
    3. A connected Azure AI Search resource used by a file search or index tool Correct answer.
    4. A second model deployment

    Explanation

    Connecting the existing Azure AI Search resource lets the agent retrieve passages from the index at query time. Additional deployments, quota increases, and content filters change capacity or safety behavior but do not provide document retrieval.

  194. Question 194Implement AI solutions by using Microsoft Foundry

    A developer must ensure that an agent never returns responses containing credit card numbers that appear in retrieved documents. Which control should be applied?

    1. Enable a content filter that detects and redacts sensitive information Correct answer.
    2. Shorten the system message
    3. Lower the model temperature
    4. Increase the chunk size of the index

    Explanation

    Sensitive-information detection in content filtering can identify and block or redact personal and financial data in prompts and responses. Temperature, chunk size, and prompt length affect generation style or retrieval granularity but do not detect regulated data patterns.

  195. Question 195Implement AI solutions by using Microsoft Foundry

    A developer wants to test an agent's behavior in the Foundry playground before publishing it. Which statement describes what the playground provides?

    1. An interactive chat surface for sending prompts and inspecting responses and tool calls Correct answer.
    2. An automated load test generator
    3. A production endpoint that serves live users
    4. A billing report for token consumption

    Explanation

    The playground is an interactive environment where developers send prompts, observe responses, and inspect tool invocations to validate behavior before deployment. It does not serve production traffic, generate load tests, or produce billing reports.

  196. Question 196Implement AI solutions by using Microsoft Foundry

    A developer notices the agent ignores retrieved documents when they are long and instead relies on its pretrained knowledge. Which change is most likely to improve grounding?

    1. Remove the system message
    2. Disable the search connection
    3. Reduce chunk size and retrieve fewer, more relevant passages Correct answer.
    4. Increase the model temperature

    Explanation

    Smaller, focused chunks with fewer but more relevant passages reduce noise and make it easier for the model to attend to the supplied evidence. Raising temperature increases randomness, and removing instructions or retrieval eliminates the grounding signal entirely.

  197. Question 197Implement AI solutions by using Microsoft Foundry

    A developer needs to give an agent a persistent memory of facts a user shared in earlier sessions, such as a preferred shipping address. Which Foundry feature should be used?

    1. A second content filter
    2. A larger context window in the model
    3. A higher temperature setting
    4. Agent memory that stores user-specific information across threads Correct answer.

    Explanation

    Agent memory persists user-specific information beyond a single conversation thread so later sessions can recall preferences. A larger context window only holds more of the current conversation, temperature affects randomness, and filters govern safety.

  198. Question 198Implement AI solutions by using Microsoft Foundry

    A developer wants to compare two model deployments for the same agent workload and determine which produces more relevant answers at acceptable cost. Which approach is most appropriate?

    1. Run the same evaluation dataset and metrics against both deployments and compare scores with token usage Correct answer.
    2. Ask the model which deployment it prefers
    3. Choose the model with the largest parameter count
    4. Select the model with the lowest latency in a single request

    Explanation

    Holding the dataset and metrics constant while varying the deployment isolates model quality differences, and pairing scores with token usage shows the cost trade-off. Parameter count, a single latency sample, or self-assessment do not measure relevance for the workload.

  199. Question 199Implement AI solutions by using Microsoft Foundry

    A developer wants to log each agent conversation so that support staff can review transcripts later. Which Foundry capability should be enabled?

    1. A second search index
    2. A higher tokens-per-minute quota
    3. A larger chunk overlap
    4. Tracing and monitoring with connected Application Insights Correct answer.

    Explanation

    Tracing and monitoring capture conversation and tool-call telemetry in Application Insights, where transcripts and traces can be reviewed. Quota, extra indexes, and chunk overlap affect capacity or retrieval but do not record conversations.

  200. Question 200Implement AI solutions by using Microsoft Foundry

    An agent occasionally calls a tool with malformed arguments, causing failures. Which adjustment is most likely to reduce these errors?

    1. Provide a precise tool description and parameter schema with examples in the instructions Correct answer.
    2. Remove the tool from the agent
    3. Reduce the model's max tokens
    4. Increase the temperature

    Explanation

    Clear tool descriptions, strict parameter schemas, and examples help the model construct valid arguments. Raising temperature increases variability and worsens malformed calls, and removing the tool abandons the capability. Reducing output tokens can truncate needed argument content.

Azure AI Fundamentals (AI-901) flashcards

31 cards. Read the term and the definition here. The deck at the top of this page is optional.

Machine learning
A branch of AI in which a model learns patterns and relationships from data rather than following explicitly programmed rules, then generalizes to new inputs.
Supervised learning
Training with labeled examples, where each input is paired with the correct answer, so the model learns to predict that answer for unseen inputs.
Unsupervised learning
Training on data without labels, letting the algorithm discover structure such as clusters or unusual outliers on its own.
Regression versus classification
Regression predicts a continuous numeric value such as price or temperature; classification predicts a discrete category such as fraud or not fraud.
Clustering
An unsupervised technique that groups items so members of a group are more similar to each other than to items in other groups, without predefined labels.
Reinforcement learning
An approach where an agent learns a policy by taking actions in an environment and receiving rewards or penalties that shape future behavior.
Object detection
A computer vision task that locates multiple objects in an image by returning bounding boxes plus a class label and confidence for each detected object.
Semantic segmentation
A vision task that classifies every pixel of an image into a category, producing a mask that shows the exact shape of each region.
Optical character recognition
A vision capability that extracts machine-readable text from images or scanned documents, converting pixels of writing into editable strings.
Natural language processing
The AI field concerned with analyzing and generating human language, including sentiment analysis, entity extraction, translation, and summarization.
Named entity recognition
An NLP task that finds and labels specific entities in text, such as people, organizations, locations, dates, and quantities.
Generative AI
Models that create new content such as text, images, or code in response to a prompt, based on patterns learned from large training corpora.
Large language model
A generative model trained on very large text collections to predict tokens, enabling chat, summarization, question answering, and code generation.
Hallucination
A generative model output that is fluent but factually unsupported or fabricated, often caused by missing grounding or ambiguous prompts.
Responsible AI principles
Six Microsoft principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability, applied across the AI lifecycle.
Anomaly detection
An unsupervised or semi-supervised technique that identifies data points or events that deviate significantly from expected or learned normal patterns.
Microsoft Foundry
Microsoft's platform for building, evaluating, and deploying AI applications and agents, organized as a resource hosting projects with models and connections.
Foundry project
A workspace inside a Foundry resource that contains model deployments, datasets, connections, agents, and evaluation runs for one application.
Model catalog
A browsable library in Foundry of foundation models from Microsoft and partners, each with descriptions, capabilities, and deployment options.
Model deployment
A hosted instance of a model that exposes an endpoint and credentials; client code targets the deployment name rather than the model family.
Standard deployment
A pay-per-token hosting option where consumption is billed by input and output tokens, with regional processing and no reserved capacity.
Provisioned throughput
A deployment option that reserves model capacity, giving predictable latency and steady cost instead of per-token billing.
Serverless API deployment
A deployment where models are hosted by Microsoft and called through an endpoint with key-based or identity-based authentication, without managing infrastructure.
System message
The prompt section that sets the model's role, tone, and constraints, applied before user messages and guiding all subsequent responses.
Few-shot prompting
Including a small number of example input-output pairs in the prompt so the model imitates the desired format and reasoning pattern.
Retrieval-augmented generation
A pattern that retrieves relevant chunks from your own data and inserts them into the prompt, grounding model responses in supplied evidence.
Azure AI Search
A cloud search service with vector and keyword indexing that is commonly used as the knowledge source for retrieval-augmented generation in Foundry.
Foundry agent
A configurable AI assistant in Foundry that combines a model deployment with instructions, tools, and knowledge sources to complete multi-step tasks.
Groundedness metric
An evaluation measure of how well a generated response is supported by the supplied source or context, rather than invented by the model.
Relevance metric
An evaluation measure of how directly a generated response addresses the user's request, independent of whether it is factually correct.
Content safety filters
Configurable controls that detect harmful categories such as violence or hate at set severity thresholds and block or flag model inputs and outputs.

Azure AI Fundamentals (AI-901) study guide

AI Concepts and Capabilities

What AI actually is

Artificial intelligence is software that performs tasks we associate with human cognition: perceiving, understanding language, reasoning, and generating content. Machine learning is the dominant approach: instead of hand-coding rules, you supply data and the system learns patterns. Deep learning uses layered neural networks and underpins most modern language and vision systems. Generative AI specifically produces new content (text, images, code, audio) rather than only classifying or predicting a label.

Workload categories to know

  • Computer vision: image classification, object detection, OCR, facial detection.
  • Natural language processing: text classification, sentiment, entity extraction, translation, summarization.
  • Speech: speech-to-text (recognition), text-to-speech (synthesis), speaker recognition.
  • Information extraction / document intelligence: pulling structured fields from forms and invoices.
  • Anomaly detection: flagging unusual data points, common in fraud and equipment monitoring.
  • Knowledge mining: indexing heterogeneous content so it becomes searchable and answerable.
  • Generative AI and agents: producing content and orchestrating multi-step tasks with tools.

Traps in questions

  1. Confusing classification (assign one label from a fixed set) with regression (predict a number).
  2. Confusing object detection (locate and label multiple objects, returns bounding boxes) with image classification (one label for the whole image).
  3. Assuming AI is always accurate or unbiased. Responsible AI principles exist precisely because models inherit bias from data and can hallucinate.
  4. Mixing up speech recognition with natural language understanding - transcription is not comprehension.
  5. Treating generative AI as a database lookup; it predicts plausible output and may fabricate.

Rule of thumb

Ask: what is the output shape? One label from a list = classification. A continuous number = regression. New free-form content = generative. Coordinates plus labels = detection. A transcript = speech recognition.

Worked example

A hospital wants to flag X-ray images that contain any abnormality (yes/no) and also mark where on the image. The yes/no decision is binary classification; the marking requires object detection. If instead the goal were estimating patient length-of-stay in days, that is regression.

Implementing AI Solutions by Using Microsoft Foundry

The platform in one paragraph

Microsoft Foundry is the unified Azure platform for building AI apps: you create a Foundry resource/project, deploy models (including models from the model catalog), and consume them through the Foundry Models inference endpoint. Foundry also provides agent building, evaluation, tracing/observability, and connections to supporting Azure services. Generative apps typically use the Azure AI OpenAI-compatible chat completions API shape.

Concepts a candidate must know

  • Deployment: a model must be deployed before you call it; the deployment has a name you reference in code, and a hosting option (for example standard pay-per-token versus provisioned throughput).
  • Tokens: billing and context limits are measured in tokens, not characters. Prompts plus completion both consume the context window.
  • Prompt engineering: system message sets behavior and constraints; user message carries the request. Few-shot examples, grounding data, and explicit output-format instructions improve reliability.
  • Retrieval-augmented generation (RAG): retrieve relevant content from your own index or search store and inject it into the prompt so answers are grounded in your data rather than model memory.
  • Agents and tools: an agent combines a model with instructions and callable tools (functions, search, code) to complete multi-step tasks.
  • Evaluation: measure groundedness, relevance, coherence, and fluency; compare prompt variants before shipping.
  • Content safety and responsible AI: filters and guardrails apply to inputs and outputs; know that configuration is part of the solution, not an afterthought.

Traps

  1. Calling a model by its catalog name when the code needs the deployment name.
  2. Assuming the model knows your private documents - without RAG or fine-tuning it does not.
  3. Forgetting that long conversations can exceed the context window; you must trim or summarize history.
  4. Confusing fine-tuning (changes model weights with your examples) with RAG (supplies context at query time).
  5. Ignoring cost drivers: output tokens usually cost more than input tokens.

Rule of thumb

If the knowledge changes often or must be cited, use RAG. If you need consistent style or task behavior across many calls, consider fine-tuning. If you need neither, a well-written system prompt is cheapest.

Worked example

A support bot must answer from a policy library updated weekly. Index the policies, retrieve the top matching passages per question, place them in the prompt, and instruct the model to answer only from that context and to say it does not know otherwise. This is RAG; re-indexing weekly keeps answers current without retraining.

Azure AI Fundamentals (AI-901) cheat sheet

Quick reference

Identify AI Concepts and Capabilities

  • Artificial intelligence is the umbrella discipline; machine learning is a subset that learns patterns from data instead of following hand-written rules.

  • Supervised learning uses labeled data and solves regression (numeric output) and classification (discrete label) problems.

  • Unsupervised learning uses unlabeled data; clustering groups similar items, and anomaly detection flags points that deviate from learned patterns.

  • Reinforcement learning trains an agent by reward and penalty signals across actions, states, and episodes rather than labeled examples.

  • Computer vision capabilities include image classification, object detection (bounding boxes), semantic segmentation (per-pixel class), and OCR text extraction.

  • Natural language processing covers text classification, sentiment, named entity recognition, key phrase extraction, translation, and summarization.

  • Generative AI produces new content (text, images, code) from prompts; large language models are generative models trained on very large text corpora.

  • Responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability.

  • Workload-to-capability rule: predicting a numeric value is regression; assigning a category is classification; grouping without labels is clustering.

Quick reference

Implement AI Solutions by Using Microsoft Foundry

  • Microsoft Foundry is the unified platform for building AI apps and agents: a Foundry resource (hub) hosts projects, and each project holds models, data, and connections.

  • The Foundry portal provides a model catalog, a playground for chat and completions, and evaluation and deployment tooling; the Foundry SDK and REST APIs automate the same operations.

  • Deployment options differ by hosting: standard (pay-per-token, regional data residency) versus provisioned throughput (reserved capacity, predictable latency) versus serverless API endpoints.

  • A model deployment exposes an endpoint and credentials; client code calls it with the project endpoint plus a deployment name, not the model family name.

  • Prompt engineering levers: system message for role and constraints, few-shot examples, explicit output format, and grounding context supplied in the prompt.

  • Retrieval-augmented generation grounds responses in your own data by retrieving relevant chunks and inserting them into the prompt, reducing unsupported answers.

  • Azure AI Search can serve as the knowledge source for RAG; Foundry agents can call tools and knowledge sources to complete multi-step tasks.

  • Evaluation metrics: groundedness, relevance, coherence, and fluency for generative output; precision, recall, and F1 for classification.

  • Content safety and responsible AI controls (harm categories, severity thresholds, prompt shields) are configured per resource and applied at inference time.

Azure AI Fundamentals (AI-901) key terms

Artificial Intelligence (AI)

The field of building computer systems that perform tasks normally requiring human intelligence, such as understanding language, recognizing images, making decisions, or generating content.

Related: Machine Learning, Generative AI
Machine Learning

A branch of AI in which systems learn patterns from data and improve with experience, rather than following rules written by hand for every case.

Related: Training Data, Model, Supervised Learning
Deep Learning

A machine learning approach that stacks many layers of artificial neurons so the system can learn complex patterns such as speech, images, and language directly from raw data.

Related: Neural Network, Foundation Model
Neural Network

A computing structure of connected nodes arranged in layers, loosely inspired by the brain, that transforms input values step by step to produce a prediction or generated output.

Related: Deep Learning, Model
Model

The trained artifact produced by a machine learning process. It holds learned parameters and turns new input data into predictions, classifications, or generated content.

Related: Training Data, Inference, Evaluation
Training Data

The labeled or unlabeled examples used to teach a model. Its quality, size, and representativeness strongly shape how accurate and fair the finished model becomes.

Related: Model, Supervised Learning, Data Labeling
Supervised Learning

A training method that uses examples paired with known correct answers, so the model learns to map inputs to expected outputs such as categories or numeric values.

Related: Classification, Regression, Data Labeling
Unsupervised Learning

A training method that works with data having no known answers, seeking hidden structure such as natural groupings or unusual records within the dataset.

Related: Clustering, Training Data
Classification

A supervised task that assigns each input to one of a fixed set of categories, for example labeling a message as spam or not spam.

Related: Supervised Learning, Regression
Regression

A supervised task that predicts a continuous numeric value, such as a price, temperature, or expected demand, instead of a discrete category.

Related: Supervised Learning, Classification
Clustering

An unsupervised technique that groups similar records together based on shared characteristics, without being told in advance what the groups should be.

Related: Unsupervised Learning
Computer Vision

The AI capability that lets systems interpret images and video, supporting tasks such as detecting objects, reading text, classifying scenes, or recognizing faces.

Related: Object Detection, Optical Character Recognition, Multimodal Model
Natural Language Processing (NLP)

The AI capability concerned with understanding and producing human language, covering tasks such as sentiment analysis, translation, summarization, and question answering.

Related: Large Language Model, Token, Named Entity Recognition
Generative AI

AI that creates new content such as text, images, audio, or code by learning patterns from large amounts of existing data and producing novel combinations on request.

Related: Large Language Model, Prompt, Foundation Model
Large Language Model (LLM)

A generative model trained on very large text collections to predict and produce language, enabling chat, summarization, translation, and code generation.

Related: Generative AI, Token, Prompt
Foundation Model

A large pretrained model that serves as a general base for many downstream tasks, typically adapted through prompting or fine-tuning rather than trained from scratch.

Related: Large Language Model, Fine-Tuning, Model Catalog
Multimodal Model

A model that accepts or produces more than one kind of content, such as combining text and images in a single request or response.

Related: Generative AI, Computer Vision
Prompt

The instruction or input text a person sends to a generative model. Its wording, context, and examples strongly influence the quality of the response.

Related: System Message, Token, Grounding
Token

A small chunk of text, roughly a word fragment, that language models process as units. Token counts drive context limits and usage-based billing.

Related: Large Language Model, Context Window
Hallucination

A confident but incorrect or fabricated output from a generative model, often occurring when the model lacks reliable information and fills gaps with plausible-sounding content.

Related: Grounding, Retrieval-Augmented Generation, Responsible AI
Responsible AI

A set of principles guiding AI development, commonly fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability, so systems treat people justly.

Related: Content Safety, Hallucination
Microsoft Foundry

Microsoft's platform for building, evaluating, deploying, and governing AI applications and agents, bringing together model access, tools, and project resources in one place.

Related: Foundry Project, Model Catalog, Deployment
Foundry Project

A workspace within Microsoft Foundry that groups related models, connections, agents, evaluations, and other resources for a single AI solution.

Related: Microsoft Foundry, Connection
Model Catalog

A browsable library of pretrained models inside Microsoft Foundry, spanning providers and modalities, from which teams can select, test, and deploy a model.

Related: Foundation Model, Deployment, Microsoft Foundry
Deployment

The act of publishing a model to an endpoint so applications can send requests to it. Settings include capacity, rate limits, and the deployment name used in calls.

Related: Endpoint, Model Catalog, Token
Endpoint

A network address and credential pair that an application calls to reach a deployed model or service, typically secured with a key or Microsoft Entra ID authentication.

Related: Deployment, Connection
Playground

An interactive area in Microsoft Foundry for trying prompts and model settings against a deployed model before writing application code around it.

Related: Prompt, System Message, Deployment
System Message

Instructional text sent ahead of user input that sets a model's role, tone, and boundaries, helping keep responses consistent across a conversation.

Related: Prompt, Temperature
Temperature

A model setting that controls randomness in generated output. Lower values make responses more predictable and focused; higher values make them more varied and creative.

Related: System Message, Prompt
Context Window

The maximum amount of text, measured in tokens, that a model can consider at once, covering the prompt, any supplied documents, and the generated reply.

Related: Token, Large Language Model
Retrieval-Augmented Generation (RAG)

A pattern that retrieves relevant content from a data source and adds it to the prompt, so answers rely on supplied facts rather than only on the model's training.

Related: Grounding, Embedding, Hallucination
Grounding

Anchoring model responses in trustworthy, supplied information such as company documents or search results, which reduces invented answers and improves relevance.

Related: Retrieval-Augmented Generation, Hallucination
Embedding

A numeric vector representing text or another item so that similar meanings sit close together, enabling semantic search, recommendations, and retrieval in RAG solutions.

Related: Retrieval-Augmented Generation, Vector Search
Vector Search

A retrieval method that finds items whose stored embeddings are most similar to a query embedding, powering semantic lookup over documents, images, or other content.

Related: Embedding, Retrieval-Augmented Generation
Fine-Tuning

Further training of an existing model on a smaller, task-specific dataset so its behavior better matches a particular domain, style, or set of examples.

Related: Foundation Model, Training Data
Agent

An AI component that pursues a goal by reasoning over a request, choosing and calling tools or knowledge sources, and returning a result, often across multiple steps.

Related: Tool, Microsoft Foundry, Prompt
Tool

A function or external service an agent or model can invoke, such as a search index, calculator, or business API, to obtain information or perform actions beyond text generation.

Related: Agent, Function Calling
Function Calling

A model capability where the model returns a structured request to run a named function with specific arguments, letting the application execute code and feed results back.

Related: Tool, Agent
Evaluation

The practice of measuring a model or application's quality using metrics and test datasets, such as groundedness, relevance, coherence, and similarity to expected answers.

Related: Model, Content Safety
Content Safety

Filtering that detects and blocks harmful categories such as hate, violence, sexual content, and self-harm in prompts and responses, with configurable severity thresholds.

Related: Responsible AI, Evaluation

Azure AI Fundamentals (AI-901) study plan

8 weeks ยท About 6 hours per week

Week 1

Identify AI concepts and capabilities

  • Define artificial intelligence, machine learning, and deep learning in your own words and explain how the three relate

  • Classify ten everyday products as examples of AI or non-AI and justify each classification

  • Describe the difference between supervised, unsupervised, and reinforcement learning with one concrete scenario for each

  • Build a personal glossary of at least 20 AI terms with a one-sentence definition for each

Your checkpoint

A one-page concept map linking AI, machine learning, deep learning, and generative AI, plus a 20-term glossary

Week 2

Identify AI concepts and capabilities

  • Identify common computer vision workloads such as image classification, object detection, and optical character recognition

  • Explain how natural language processing tasks differ, including text classification, entity recognition, and sentiment analysis

  • Distinguish generative AI workloads from traditional predictive workloads and state when each is appropriate

  • Summarize the responsibilities in a responsible AI framework and match each principle to a realistic failure scenario

Your checkpoint

A workload-matching table listing 12 scenarios, the correct AI workload category, and a one-line rationale for each

Week 3

Identify AI concepts and capabilities

  • Describe how a model is trained, evaluated, and deployed across the machine learning lifecycle

  • Explain the purpose of training data, validation data, and test data splits

  • Interpret common evaluation metrics such as accuracy, precision, recall, and F1 score for a classification problem

  • Identify the ethical, privacy, and bias risks that appear at each stage of the lifecycle

Your checkpoint

A written walkthrough of one end-to-end machine learning scenario covering data, training, evaluation, and deployment in 400 or more words

Week 4

Implement AI solutions by using Microsoft Foundry

  • Navigate the Microsoft Foundry portal and locate model catalog, playground, and project resources

  • Create a project and deploy a model from the catalog by using the portal interface

  • Explain the roles of a hub, a project, and a deployment in the Foundry resource hierarchy

  • Compare at least three available models on cost, capability, and intended use

Your checkpoint

A working Foundry project with one deployed model and a screenshot-free written record of the resource names and configuration choices

Week 5

Implement AI solutions by using Microsoft Foundry

  • Write effective system messages and prompts that constrain model behavior for a defined task

  • Apply prompt engineering techniques such as few-shot examples and explicit output formatting

  • Test a single prompt against three model configurations and record differences in output quality

  • Identify signs of hallucination in generated output and describe a mitigation for each sign

Your checkpoint

A prompt library of five prompts, each with the system message, expected output format, and observed results

Week 6

Implement AI solutions by using Microsoft Foundry

  • Explain how grounding with your own data improves answer relevance and reduces unsupported claims

  • Describe the purpose of embeddings and vector search in a retrieval-augmented generation pattern

  • Configure a Foundry solution that answers questions from a provided document set

  • Evaluate grounded responses for faithfulness to the source documents and note failure cases

Your checkpoint

A grounded question-and-answer solution over a small document set, with ten test questions and recorded answers

Week 7

Implement AI solutions by using Microsoft Foundry

  • Identify content safety, prompt injection, and data leakage risks in a generative AI solution

  • Describe how responsible AI controls such as content filters and evaluation metrics are applied in Foundry

  • Run an evaluation of a deployed model and interpret the reported quality and safety metrics

  • Explain how monitoring and logging support ongoing operation of an AI solution

Your checkpoint

An evaluation report summarizing quality and safety metrics for one deployed model, with three recommended improvements

Week 8

Identify AI concepts and capabilities & Implement AI solutions by using Microsoft Foundry

  • Review weak areas identified in earlier deliverables and correct at least five misconceptions

  • Explain each Foundry implementation step in terms of the underlying AI concept it applies

  • Complete two timed practice sets and analyze every missed item by domain

  • State the key differences between AI concepts and Foundry implementation tasks in a consolidated summary

Your checkpoint

A final one-page summary sheet plus a scored practice log showing per-domain accuracy above 80 percent

Milestones

Week 2: Score at least 80 percent on a 20-question self-quiz covering AI concepts, workloads, and responsible AI principles

Week 4: Demonstrate a deployed model in a Foundry project and explain the hub, project, and deployment hierarchy without notes

Week 6: Show a grounded solution answering questions from your own documents, with at least 8 of 10 test answers judged faithful to the sources

Week 8: Achieve at least 80 percent on two timed practice sets, with no domain below 70 percent

Azure AI Fundamentals (AI-901) exam day

Your checklist

Government-issued photo ID

Bring the original physical ID (passport or government ID) whose name matches your registration exactly. Confirm with the provider which ID types are accepted for your location, and whether a secondary ID is required; do not rely on a photocopy or a digital ID unless the provider confirms it is accepted.

Appointment confirmation and registration details

Have your appointment confirmation, registration ID, and the email/phone used at booking available. You may be asked for the exam name and appointment time at check-in.

Test-center logistics confirmed

Confirm the exact address, building entrance, and arrival time with the provider the day before. Arrive early to allow for security checks and check-in; late arrival can forfeit your seat.

Online-proctoring system test completed

If you are testing online, run the provider's system/pre-check test in advance on the same computer and network you will use. Confirm with the provider the supported browser, operating system, webcam, microphone, and bandwidth requirements.

Quiet, private, well-lit room (online delivery)

Use a room with a closed door, no other people, and a clear desk. Remove notes, phones, watches, and extra screens. Confirm with the provider what is allowed on the desk and whether a room scan is required.

Know your login/launch procedure

Know how you launch the exam (provider portal, emailed link, or proctor check-in). Write down the steps and support phone number so you are not searching during check-in.

Study summary of the AI-901 skill areas

Review a one-page summary of the official skills outline: AI workloads and considerations, machine learning fundamentals, computer vision, natural language processing, and generative AI on Azure, plus responsible AI principles. Verify the current outline and weighting on the provider's official page.

Practice with the question format

Be comfortable with multiple-choice, multi-select, drag-and-drop, and hotspot-style items if the provider lists them. Read each item fully and note whether it asks for one answer or several. Confirm the current item types with the provider.

Scratch paper and pen policy confirmed

Many providers supply or allow an erasable note board and pen, or an online whiteboard. Do not assume; confirm with the provider what is permitted and whether you must supply it yourself.

Break and timing plan

Know the total appointment time versus the exam time and whether breaks are offered. Confirm with the provider whether breaks are available and whether the clock stops during them.

Plan for technical failure

Know the provider's support contact and their procedure if the connection drops or the exam is interrupted. Save the support number offline in case your internet fails.

Accessibility or accommodation requests confirmed

If you need extra time or other accommodations, confirm in advance that your request was approved. Accommodation policies and lead times vary; confirm with the provider.

Plan the day

Final 72 hours

Stop adding new material. Work through the official skills outline and mark any objective you cannot explain in one or two sentences. Review those gaps using your own notes and the official learning paths. Do one timed practice set to check pacing, then review wrong answers for the concept, not the wording.

Final 72 hours

Write a one-page cheat sheet in your own words: what each Azure AI service category is for, the difference between prebuilt and custom models, when to use classification versus regression versus clustering, and the responsible AI principles. Rewriting forces recall and gives you a last-minute review sheet.

Final 72 hours

Confirm logistics: appointment time, test-center address or online check-in steps, ID requirements, and allowed materials. Confirm with the provider anything you are unsure about rather than guessing on exam day.

Final 24 hours

Do a light review of your cheat sheet and the skills outline only. No new topics and no full-length practice tests; late cramming tends to increase anxiety without improving recall.

Final 24 hours

Prepare and lay out your ID, confirmation details, and any permitted items. For online delivery, run the system test once more, update the browser if needed, and charge your laptop. Check the room setup and tell household members your exam window.

Final 24 hours

Protect sleep. Set two alarms, plan the route or the network backup (for example, a mobile hotspot), and decide what you will do if the connection fails. Aim to arrive or log in early rather than exactly on time.

Morning of

Eat a normal meal and hydrate, but avoid heavy caffeine if you are already anxious. Do a 15-20 minute review of your cheat sheet only. Do not attempt new practice questions.

Morning of

Recheck the appointment details and ID. For a test center, leave with buffer time for traffic and security. For online delivery, close background apps, silence notifications, and keep your phone out of reach unless the proctor instructs otherwise.

During the test

Read the full item before looking at the options, and watch for words like 'not', 'first', or 'best'. For multi-select items, count how many answers are requested. If an item is taking too long, flag it, choose your best answer, and move on.

During the test

Answer every item; there is normally no penalty for guessing, but confirm the scoring rules with the provider. Use remaining time to revisit flagged items, and change an answer only when you can point to a specific reason.

During the test

If a technical problem occurs, stop and contact the proctor or support immediately rather than troubleshooting alone, and note the time. Follow the provider's incident procedure so the interruption is documented.

Rules to check

  • Bring the exact ID required and make sure the name matches your registration. If your ID name differs from your registration, resolve it with the provider before exam day; do not assume staff can change it at check-in.

  • Follow all proctor instructions. Do not talk, do not let others enter the room, and do not use phones, smart watches, notes, or a second screen unless the provider explicitly permits them.

  • Do not attempt to capture, copy, or share exam content. Most providers require you to accept a non-disclosure agreement before the exam starts; violating it can void your result and lead to sanctions.

  • Understand the scoring before you sit: scaled scores, pass marks, and whether unanswered items count against you vary by exam. Confirm the current pass score and scoring method with the provider rather than relying on hearsay.

  • Know the reschedule and cancellation policy in advance. Deadlines and fees vary by provider and location, so confirm with the provider if you may need to move your appointment.

  • If you test online, keep the room private for the whole session and follow the room-scan and desk-clear instructions. Leaving the camera view or the room may end the session.

  • Report any technical interruption to the proctor or support immediately and follow their instructions; do not restart the exam on your own unless told to do so.

  • Results and score-report delivery timing vary. Confirm with the provider how and when your result will be available and what the retake policy is.

Azure AI Fundamentals (AI-901) FAQ

How many Azure AI Fundamentals (AI-901) practice questions are available?

The practice set has 200 Azure AI Fundamentals (AI-901) questions. Each one includes the choices, the correct answer, and an explanation, and the set is free to read with no signup.

How many Azure AI Fundamentals (AI-901) flashcards are available?

The flashcard set has 31 Azure AI Fundamentals (AI-901) cards. Each card shows the term and its definition.

Which topics does the Azure AI Fundamentals (AI-901) study material cover?

The Azure AI Fundamentals (AI-901) study material covers AI Concepts and Capabilities and Implementing AI Solutions by Using Microsoft Foundry.