Microsoft AI-103 Practice Exam

• A. A text-only language model with a large context window
• B. A multimodal model that accepts image input Correct answer
• C. A speech model that transcribes spoken questions
• D. A text-embedding model for product descriptions

Option B is correct because answering questions about what is visible in a photo requires a multimodal model that accepts image input. Option A cannot see images, option C only handles speech, and option D produces vectors for similarity search rather than answers.

• A. Vector search over chunk embeddings
• B. Hybrid search that combines keyword and vector queries Correct answer
• C. Full-text search with a custom analyzer
• D. Filtering on a policy code field with OData expressions

Option B is correct because hybrid search runs keyword and vector queries together and merges the results, so exact codes like HR-2291 match by keyword while topical descriptions match by semantic similarity. Vector-only search (A) can miss exact codes, while full-text (C) and OData filters (D) do not handle natural-language topic queries well.

• A. Keep the fine-tuned model and paste each week's regulation text into the system message
• B. Retrain the fine-tuned model every week on the latest regulation text
• C. Switch to the base model and write the firm's style rules into the system message
• D. Keep the fine-tuned model and ground it on an Azure AI Search index of the regulations Correct answer

Option D is correct because grounding the model on an Azure AI Search index keeps the learned writing style, always uses the current regulation text, and lets answers cite their sources. Pasting text into the system message (A) does not scale, weekly retraining (B) is slow and still gives no citations, and dropping the fine-tune (C) loses the style and the regulation content.

• A. Set a tokens-per-minute limit on the deployment Correct answer
• B. Create a budget with an alert in Microsoft Cost Management Correct answer
• C. Assign the Cost Management Reader role to the finance team
• D. Send deployment logs to Log Analytics through diagnostic settings
• E. Configure autoscale rules on the deployment

Options A and B are correct because a tokens-per-minute limit on the deployment caps how fast the app can consume tokens, and a Microsoft Cost Management budget with an alert emails stakeholders as monthly spend approaches the planned amount. The Cost Management Reader role only grants visibility, diagnostic logs to Log Analytics record usage without limiting it or alerting on cost, and autoscale rules do not apply to a pay-as-you-go model deployment.

• A. The name of the Foundry project
• B. The deployment name Correct answer
• C. The model name shown in the model catalog
• D. The endpoint URL of the deployment

Option B is correct because with the OpenAI-compatible client the model parameter carries the name you gave the deployment, not the underlying model name. The project name, catalog model name and endpoint URL do not identify which deployment to route to.

• A. Retrieve passages from an Azure AI Search index of the PDFs Correct answer
• B. Add sample policy question-and-answer pairs as few-shot examples
• C. Fine-tune the deployed model nightly on the newest PDFs
• D. Paste the entire policy library into the system message

Option A is correct because retrieval-augmented generation over an Azure AI Search index can be re-indexed as new PDFs are published, so answers follow the newest files within a day. Few-shot examples and nightly fine-tuning do not reliably carry daily-changing facts, and pasting the whole library into the system message exceeds practical context limits.

• A. Groundedness Correct answer
• B. Coherence
• C. Fluency
• D. Relevance

Option A is correct because the Groundedness evaluator measures whether the response is supported by the retrieved context. Coherence, fluency and relevance judge readability, flow and topical fit, and none of them detects unsupported claims.

• A. Provision Azure resources by deploying Bicep templates
• B. Run Python code in a sandboxed session Correct answer
• C. Call a private REST API from inside the sandbox
• D. Search current web pages for live content
• E. Read data from files attached to the agent Correct answer
• F. Create chart images and other output files Correct answer

Options B, E and F are correct because the code interpreter tool runs Python in a sandboxed session, can read files attached to the agent, and can generate output files such as charts. Option A is wrong because the sandbox does not deploy Azure resources, option C is wrong because it has no route to private REST APIs (use OpenAPI or function tools), and option D is wrong because live web search belongs to the web search tool.

• A. Set n to 4 in the image generation request Correct answer
• B. Set the size value to a square resolution
• C. Set the quality value to high in the request
• D. Set the output compression to a lower level

Option A is correct because the n parameter sets how many images gpt-image-1 returns for one prompt, so n=4 gives four alternatives in a single request. Size, quality and output compression change the look or file size of each image, not how many are returned.

• A. A size value matching the page layout so the background fits the page
• B. A JPEG output format with the highest quality value to keep edges clean
• C. A background value of transparent with PNG or WebP as the output format Correct answer
• D. A prompt asking for a white background, then converting white pixels to alpha later

Option C is correct because gpt-image-1 supports a background value of transparent, which only works with an output format that has an alpha channel, such as PNG or WebP. JPEG has no transparency, and size or a prompt for a white background cannot produce true alpha.

• A. Read it from the default container of the project storage account
• B. Repeat the call with a longer expiry value so the link stays valid
• C. Poll the same request until the moderation step adds a URL
• D. Decode the base64 data that the response returns for the image Correct answer

Option D is correct because gpt-image-1 always returns the image as base64-encoded data (b64_json) rather than a hosted URL, so the app decodes that data to get the picture. There is no expiring link to extend, no storage container written for you, and nothing to poll.

• A. Set stream to true on the image request Correct answer
• B. Set background to transparent so early layers display first
• C. Set quality to low so the render finishes sooner
• D. Set partial_images to the number of previews wanted Correct answer
• E. Set n to a larger value so extra pictures arrive early

Options A and D are correct because previews arrive only when streaming is turned on and partial_images says how many in-progress images to send. Option C only makes the render faster, option B controls transparency, and option E asks for more final pictures rather than early previews.

• A. Key phrase extraction Correct answer
• B. Conversation summarization
• C. Sentiment analysis
• D. Named entity recognition

Option A is correct because key phrase extraction returns the main talking points of each ticket, such as password reset, with no training. Sentiment analysis scores tone, named entity recognition finds categorized items like people or places, and conversation summarization condenses dialogues.

• A. Conversational language understanding
• B. Key phrase extraction
• C. Language detection Correct answer
• D. Entity linking

Option C is correct because language detection returns the language name, ISO code and a confidence score for each document, which is what the routing needs. Conversational language understanding finds intents, key phrase extraction finds topics, and entity linking ties mentions to knowledge-base entries.

• A. Turn on the opinion mining option in the sentiment analysis request Correct answer
• B. Send every sentence as its own document and compare the sentence scores
• C. Run entity recognition and read the sentiment label attached to each entity
• D. Run key phrase extraction, then score each phrase with a separate sentiment call

Option A is correct because opinion mining, enabled in the sentiment analysis request, links each opinion to the aspect it describes, such as the room or the staff. Sentence scores, entity-level labels and per-phrase scoring do not return aspect-level opinions.

• A. The emotion the author feels, such as anger or joy
• B. A label with confidence scores for each sentence Correct answer
• C. An overall label of positive, neutral, negative or mixed Correct answer
• D. A star rating from one to five for the review
• E. The intent behind the review, such as complaint or praise

Options B and C are correct because sentiment analysis returns an overall document label of positive, neutral, negative or mixed, and a label with confidence scores for each sentence. It does not detect specific emotions such as anger or joy, assign star ratings, or identify the author's intent.

• A. A soft-delete deletion detection policy on a blob metadata property Correct answer
• B. A high water mark change detection policy on each blob's last-modified timestamp
• C. A field mapping that copies the blob path into the document key field
• D. A shorter indexer schedule so that each run scans the whole container more often

Option A is correct because a soft-delete deletion detection policy tells the indexer to remove a document from the index when a chosen blob metadata property holds the delete marker value. A high water mark policy (B) only detects changed blobs, a field mapping on the key (C) does not detect deletions, and scanning more often (D) still never sees a marked blob as deleted.

• A. Enable the enrichment cache on the indexer with a storage connection Correct answer
• B. Add a knowledge store with projections so later runs read the enriched output from storage
• C. Lower the indexer batch size so that fewer documents reach the skillset in each run
• D. Reset the indexer and rerun it so that every skill regenerates its output from the source files

Option A is correct because the enrichment cache stores the output of each skill, so after only the embedding skill changes, the OCR and image-description results are reused instead of being billed again. A knowledge store (B) is not read back by the indexer, a smaller batch size (C) does not avoid repeated calls, and a reset (D) forces every skill to run again.

• A. An index projection that fans each document out into child search documents
• B. A knowledge store definition with table projections Correct answer
• C. A debug session that saves the enrichment tree to a storage container
• D. An output field mapping that writes the enriched fields into a second search index

Option B is correct because a knowledge store with table projections saves enriched content as tables in Azure Storage, outside the search index, where Power BI can load it. Index projections (A) and a second index through output field mappings (D) write to search indexes, and a debug session (C) is a troubleshooting tool, not a durable output.

• A. facet on a category field to return a count per value
• B. $select limited to the title field to shrink each result
• C. $top set to the page size and $skip set to the offset of the page Correct answer
• D. searchMode set to all to require every term in each result
• E. $count set to true to return the total number of matches Correct answer

Options C and E are correct because $top with $skip returns one page at an offset, and $count set to true adds the total number of matches to the response. A facet counts values per category rather than the total, $select only shrinks each result, and searchMode all changes which documents match, which is already correct.