AI-901 AI models and configuration practice questions

12 sample questions with full explanations. Try each one before you open the answer. New to this area? Read the AI models and configuration study guide first.

Question 1

A user asks a chat model a question, and the answer appears a few words at a time. Which statement best describes how the model is producing the answer?
- A It looks up a stored answer to the same question that it memorized from its training data
- B It predicts the next token, adds it, and repeats
- C It searches the web for each word before writing it
- D It writes the whole answer at once, and the app adds a delay to make it look typed
Show answer and explanation
Answer: B

Language models generate completions one token at a time. Each predicted token is added to the sequence, and the model then predicts the next one, until it predicts that the sequence has ended. That's why answers stream in piece by piece.
- A: Models don't store and retrieve answers. They generate new text from the relationships between tokens learned in training.
- C: A model can be given a web search tool, but generating text doesn't involve searching for each word.
- D: Streaming reflects the real token-by-token generation, not a delay added by the app.
Source: Large language models (LLMs) (Microsoft Learn)
Question 2

A store's search app compares the embedding of a shopper's query with embeddings of product descriptions. A search for "puppy food" returns a product described as "dog food", even though the word "puppy" isn't in its description. Why does this work?
- A Embeddings for words used in similar contexts point in similar directions, so their vectors are close
- B The model converts "puppy" and "dog" into the same token ID
- C The app stores a hand-written list of synonyms for every product
- D Embeddings match words that share the same first few letters
Show answer and explanation
Answer: A

An embedding is a vector that captures a token's meaning based on the contexts it appears in. Words with similar meanings, like "dog" and "puppy", end up with vectors pointing in similar directions, so semantic search can match them even without the same words.
- B: Each distinct token has its own ID. Similarity comes from the embedding vectors, not shared IDs.
- C: A synonym list would be keyword matching. Embeddings capture similar meaning without one.
- D: Embeddings are based on meaning, not spelling. "Dog" and "puppy" share no letters.
Source: Large language models (LLMs) (Microsoft Learn)

Question 3

A team is deciding between a large language model (LLM) and a small language model (SLM). For each statement, choose Yes if it is true or No if it is false.

An SLM can be a good fit for an assistant that must run on a laptop or phone instead of in the cloud.
LLMs generalize well but can cost more to train and use.
An SLM always gives more accurate answers than an LLM.

Show answer and explanation

An SLM can be a good fit for an assistant that must run on a laptop or phone instead of in the cloud.	Yes
LLMs generalize well but can cost more to train and use.	Yes
An SLM always gives more accurate answers than an LLM.	No

LLMs are powerful and handle a wide range of tasks, but they cost more to run. SLMs work well for focused topics and for small, easily deployed models on devices. Neither is always more accurate; the right choice depends on the task.

Source: Generative AI and agents (Introduction to AI concepts module) (Microsoft Learn)

Question 4

You're building a tutoring app on a chat model. Whatever the student asks, every answer must use a friendly tone, explain step by step, and decline medical questions. Where should you put these rules?
- A In each user prompt
- B In a stop sequence
- C In the system prompt
- D In the max output tokens setting
Show answer and explanation
Answer: C

The system prompt sets the assistant's behavior, tone and guardrails for the whole conversation. The user prompt is the student's actual question, which changes every time.
- A: The user prompt holds the end user's request. Rules that must always apply belong in the system prompt, where users can't forget or change them.
- B: A stop sequence only tells the model where to stop generating text.
- D: Max output tokens only limits response length.
Source: Using a generative AI model (Get started with generative AI and agents in Microsoft Foundry) (Microsoft Learn)
Question 5

You need to turn thousands of product descriptions into vectors so that an app can find products with similar meanings. Which type of model should you deploy from the Foundry model catalog?
- A A small language model
- B An embedding model
- C A multimodal chat model
- D A text analysis tool that extracts key phrases
Show answer and explanation
Answer: B

Embedding models are built to produce vector representations of text that capture meaning. Those vectors are what semantic search compares.
- A: A small language model generates text. It isn't designed to output vectors for search.
- C: A chat model generates text responses. It isn't designed to output vectors for search.
- D: Key phrase extraction returns important words and phrases, not vectors that capture overall meaning, so it can't find similar products that use different words.
Source: Generative AI models (Get started with generative AI and agents in Microsoft Foundry) (Microsoft Learn)
Question 6

An insurance app lets customers upload photos of a damaged car and ask questions such as "Is the bumper cracked?". Which type of model should the app use to answer?
- A A text-only small language model
- B An embedding model
- C An image generation model
- D A multimodal model
Show answer and explanation
Answer: D

Multimodal models accept images (and in some cases audio or video) alongside text in a chat request, so they can answer questions about a photo.
- A: A text-only model can't look at the photo.
- B: An embedding model turns content into vectors. It doesn't answer questions.
- C: An image generation model creates new images. It doesn't analyze photos that users upload.
Source: Generative AI models (Get started with generative AI and agents in Microsoft Foundry) (Microsoft Learn)
Question 7

A field-inspection app must suggest repair steps on a phone, even where there's no internet connection. Which kind of model is the best fit?
- A A small language model that can run on the device
- B A large frontier model in a Global Standard deployment
- C A large language model packaged to run on the phone
- D A multimodal model deployed in Microsoft Foundry
Show answer and explanation
Answer: A

Small language models are compact enough to deploy locally, so they suit lightweight phone apps that must work offline. Microsoft's example is exactly this: a lightweight phone app may run best on a small language model.
- B: A Global Standard deployment runs in Azure and needs an internet connection.
- C: Large language models are powerful but costly to run, and generally too big for a phone. Small language models are the ones suited to running on devices.
- D: A model deployed in Foundry is called over the internet, so it doesn't work without a connection.
Source: Generative AI models (Get started with generative AI and agents in Microsoft Foundry) (Microsoft Learn)

Question 8

A team uses evaluators in Microsoft Foundry to compare two models for a customer-support app. For each statement, choose Yes if it is true or No if it is false.

Groundedness is an example of an AI-assisted evaluation metric.
Foundry evaluators automatically fix the problems they find in a model's responses.
Evaluators can help the team decide which model gives better answers for their use case.

Show answer and explanation

Groundedness is an example of an AI-assisted evaluation metric.	Yes
Foundry evaluators automatically fix the problems they find in a model's responses.	No
Evaluators can help the team decide which model gives better answers for their use case.	Yes

Evaluators score responses on metrics such as groundedness, relevance, coherence and fluency, which helps you compare models. They detect and score issues, but they don't resolve them; fixing problems is up to you.

Source: Generative AI models (Get started with generative AI and agents in Microsoft Foundry) (Microsoft Learn)

Question 9

A company based in Germany must keep all model inference processing inside the EU. It wants pay-per-token billing and higher default quotas than a single-geography deployment offers. Which deployment type should it choose?
- A Global Standard
- B Global Batch
- C Data Zone Standard
- D Global Provisioned
Show answer and explanation
Answer: C

Data Zone Standard routes traffic only to datacenters within a Microsoft-defined data zone (US, EU or APAC), bills per token, and has higher default quotas than geography-based deployment types.
- A: Global Standard may process data in any Azure region, so it can't guarantee EU-only processing.
- B: Global Batch is for large asynchronous jobs and also processes data globally.
- D: Global Provisioned reserves capacity, which isn't pay-per-token, and processes data globally.
Source: Deployment types for Microsoft Foundry Models (Microsoft Learn)

Question 10

You deploy a chat model in Microsoft Foundry and assign it a Tokens Per Minute (TPM) allocation. For each statement, choose Yes if it is true or No if it is false.

The TPM allocation limits how many tokens the deployment can process each minute.
Long prompts use up the TPM allocation faster than short prompts.
Only the tokens in the model's responses count toward the TPM allocation.

Show answer and explanation

The TPM allocation limits how many tokens the deployment can process each minute.	Yes
Long prompts use up the TPM allocation faster than short prompts.	Yes
Only the tokens in the model's responses count toward the TPM allocation.	No

TPM caps how many tokens a deployment processes per minute. Both the tokens you send in prompts and the tokens the model generates count, so long prompts and long responses use the allocation faster.

Source: Generative AI models (Get started with generative AI and agents in Microsoft Foundry) (Microsoft Learn)

Question 11

You lower the temperature of a chat deployment to get more consistent answers. A colleague suggests also lowering top_p in the same request. What does Microsoft's guidance recommend?
- A Adjust temperature or top_p, but not both in the same request
- B Always change temperature and top_p together
- C Set top_p to the same value as temperature
- D Ignore top_p, because it only affects response length
Show answer and explanation
Answer: A

Temperature and top_p both control how random the output is. Microsoft recommends changing one or the other, because the effect of changing both is hard to predict.
- B: Changing both together is exactly what the guidance advises against.
- C: There's no rule that the two values should match.
- D: top_p controls which tokens the model considers (nucleus sampling). Response length is controlled by max tokens.
Source: Get Chat Completions (REST API reference) (Microsoft Learn)
Question 12

You deploy a chat model in Microsoft Foundry. Users report that it gives very different answers each time they send the same prompt, and you want responses to be more focused and consistent. Which parameter should you lower?
- A Temperature
- B Max tokens
- C Frequency penalty
- D Stop sequences
Show answer and explanation
Answer: A

Temperature controls how random the model's output is. Lowering it makes responses more focused and deterministic, so the same prompt produces more similar answers.
- B: Max tokens only limits how long a response can be.
- C: Frequency penalty discourages the model from repeating the same words. It doesn't make answers more consistent between requests.
- D: Stop sequences tell the model where to end its output.
Source: Get Chat Completions (REST API reference) (Microsoft Learn)

Want more practice on AI models and configuration?

Practice the full question bank for this area, in random order, with your progress saved in your browser.

Practice AI models and configuration