AI-103

Free AI-103 practice questions

10 questions from our AI-103 practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.

Back to the AI-103 practice exams

Question 1 of 10, Plan and manage an Azure AI solution

An agent built with Foundry Agent Service must store conversation state and files only in resources the company owns, and all traffic must stay on private networking. Which setup should you choose?

  1. AThe standard agent setup with public network access left on so that the hosted tools can reach your resources
  2. BThe basic agent setup combined with a customer-managed key in Key Vault to encrypt the Microsoft-managed storage
  3. CThe basic agent setup, which keeps state in Microsoft-managed storage
  4. DThe standard agent setup with your own Cosmos DB, Storage, and AI Search in a delegated subnet Correct answer

Answer and explanation: The standard setup lets you bring your own Cosmos DB, Storage, and Azure AI Search so that agent state lives in your subscription, and network-secured variants inject the agent runtime into your virtual network. The basic setup uses Microsoft-managed storage, which fails the ownership requirement, and adding a key does not change where data lives. Leaving public access on violates the private networking requirement.

Question 2 of 10, Plan and manage an Azure AI solution

A company is designing a Foundry agent that must remember each customer's stated preferences across separate conversations held weeks apart. Which capability should you plan for?

  1. ALong-term agent memory that persists user facts across conversations and recalls them later Correct answer
  2. BA higher temperature so that the model generalizes from older sessions
  3. CA larger context window on the model so that earlier conversations stay inside the prompt between sessions
  4. DA conversation thread that is kept open indefinitely so that each new session appends to the same history

Answer and explanation: Preferences that must survive across separate sessions belong in a persistent memory store that the agent writes to and retrieves from, independent of a single conversation. A larger context window only helps within one request and does not carry data between sessions. Keeping a single thread open forever grows token cost and still mixes unrelated sessions, and temperature only changes sampling randomness.

Question 3 of 10, Plan and manage an Azure AI solution

You are building a retrieval-augmented generation (RAG) solution over 2 million product manuals. Users ask natural-language questions that mix exact part numbers with descriptive phrases. Which retrieval approach should you choose in Azure AI Search?

  1. AHybrid search that runs keyword and vector queries together and merges results with Reciprocal Rank Fusion Correct answer
  2. BPure full-text keyword search with a custom analyzer, because descriptive phrases are tokenized just as well
  3. CFilter-only queries on a part-number field, then pass all of the matching manuals to the model as context
  4. DPure vector search over embeddings, because embeddings capture part numbers as reliably as plain keywords

Answer and explanation: Hybrid search combines BM25 keyword matching, which is strong for exact tokens such as part numbers, with vector similarity, which handles paraphrased descriptions, and fuses the two rankings with RRF. Pure vector search often misses exact identifiers, while pure keyword search misses semantically similar wording. Filtering alone does not rank by relevance and would flood the prompt with whole manuals.

Question 4 of 10, Implement generative AI and agentic solutions

A company already maintains a curated Azure AI Search index with semantic configuration and security filters. An agent must ground answers on that index. Which tool should you attach?

  1. AFile Search, after exporting every indexed document and uploading it to a new vector store
  2. BGrounding with Bing Search, scoped to the company's public website domain name
  3. CCode Interpreter, with a script that downloads the index content at the start of each run
  4. DThe Azure AI Search tool, connected to the existing index through a project connection Correct answer

Answer and explanation: The Azure AI Search tool lets an agent query an existing index through a project connection, reusing its semantic configuration and filters. Re-uploading content to File Search duplicates data and loses the curated index features. Bing grounding searches the public web, and downloading the index in Code Interpreter is inefficient and bypasses search ranking.

Question 5 of 10, Implement generative AI and agentic solutions

A Python app must connect to a Foundry project and obtain an OpenAI-compatible client for a model deployed in that project. Which object from the Azure AI Projects SDK should you create first?

  1. ATextAnalyticsClient, initialized with the Language endpoint and an Entra ID credential
  2. BAIProjectClient, initialized with the project endpoint and a DefaultAzureCredential Correct answer
  3. CBlobServiceClient, initialized with the storage account URL of the project
  4. DSearchClient, initialized with the search endpoint and an admin key

Answer and explanation: AIProjectClient from the azure-ai-projects package connects to a project by its endpoint and can return an authenticated OpenAI client and access agents, connections, and evaluations. SearchClient queries an Azure AI Search index, TextAnalyticsClient calls Language features, and BlobServiceClient reads storage; none of them represents the Foundry project.

Question 6 of 10, Implement generative AI and agentic solutions

A loan-processing workflow has fixed steps: extract data, check credit, request a manager's approval for amounts above a threshold, then notify the applicant. Steps must always run in this order with an auditable path. Which approach is MOST appropriate?

  1. AA group-chat orchestration where several agents debate the next step until they reach consensus
  2. BA deterministic workflow of agents and actions with a human approval step for large loans Correct answer
  3. CA single autonomous agent with all tools enabled that decides the order of steps on every request
  4. DA connected-agent setup where the credit agent decides independently whether approval is needed

Answer and explanation: When the business process is fixed and must be auditable, a defined workflow that orchestrates agents and actions in sequence, with an explicit human-in-the-loop step, is the right choice. Letting a single autonomous agent or a debating group choose the order makes the path nondeterministic. Leaving the approval decision to an agent undermines the control.

Question 7 of 10, Implement computer vision solutions

A training company has hour-long recorded lectures. It needs chapter segments with a title, a summary, and key topics for each segment to power navigation. Which solution requires the LEAST custom code?

  1. AUpload the video to Blob Storage and enable a full-text index on the video file metadata
  2. BTranscribe the audio with batch transcription and split the transcript every ten minutes
  3. CExtract one frame per second, caption each frame with a chat model, and cluster captions by hand
  4. DA Content Understanding video analyzer that segments the video and generates fields per segment Correct answer

Answer and explanation: Content Understanding video analysis can segment content and produce per-segment fields such as titles and summaries from both visual and audio signals. Frame-by-frame captioning and clustering is heavy custom work. Fixed-interval transcript splits ignore topic boundaries, and indexing file metadata reveals nothing about the content.

Question 8 of 10, Implement computer vision solutions

A video archive analysis must return, for each scene, the people count, the setting, and a summary, while also capturing what is said. Which analyzer input does Content Understanding use to produce this?

  1. AOnly the video file name and its creation date metadata
  2. BBoth the visual frames and the audio transcript of the video Correct answer
  3. COnly the audio track converted to text, with no visual input used
  4. DOnly one frame per video chosen at the midpoint of the file

Answer and explanation: Video analysis combines visual frames with the transcript so that fields can reflect both what is seen and what is said. File metadata carries no scene content, audio alone misses visual facts such as people count, and one frame cannot describe every scene.

Question 9 of 10, Implement text analysis solutions

A multinational support agent must translate customer messages, preserve product names, and adapt informal slang into polite business language in the target language. Which approach BEST satisfies all requirements?

  1. AAzure Translator text translation with the default general model and no further settings
  2. BAn LLM-powered translation prompt with a glossary of product names and tone instructions Correct answer
  3. CTransliteration of each message into the agent's script followed by manual editing
  4. DSpeech translation with the message read aloud by text to speech before translation

Answer and explanation: An LLM translation flow can follow a glossary and tone instructions to rephrase slang into polite language while preserving terms. The default Translator model translates faithfully but does not rewrite tone by instruction. Transliteration does not translate, and adding speech steps changes nothing about tone or terminology.

Question 10 of 10, Implement information extraction solutions. Choose 2.

Scanned invoices in Blob Storage contain no embedded text. Text from these images must be searchable in Azure AI Search. Which two settings are required? (Choose two.)

  1. AAdd the OCR skill to the skillset and map its text output to an index field Correct answer
  2. BAdd the Language Detection skill to infer the script of each scanned page
  3. CSet the content field's analyzer to the standard Lucene analyzer
  4. DConfigure the indexer to generate normalized images from the files Correct answer
  5. EEnable the semantic ranker on the index to read text inside images

Answer and explanation: The indexer must extract images (imageAction set to generate normalized images) so that the OCR skill can read text from them, and the OCR output must be mapped into the index. Language detection works on text that already exists. The semantic ranker reranks text results and cannot read images, and analyzers tokenize text that has already been extracted.