Task 1.2: Identify AI model components and configurations

• AIt repeatedly predicts the most likely next token from the text so far and appends it until the response is complete

  Correct

  Language models are trained to complete a sequence. Each predicted token is appended and the process repeats, which is how the whole reply is built.

• BIt finds the stored question most similar to the prompt and returns the answer that was saved with that question

  Incorrect

  Matching a question to a stored answer is how a question-and-answer knowledge base works. A language model writes new text instead of returning saved answers.

• CIt sorts the prompt into one of a fixed set of intents and fills in the response template written for that intent

  Incorrect

  Intent classification with templated replies is how many predictable conversational bots work. A language model isn't limited to a fixed set of intents or templates.

• DIt applies a set of hand-written grammar rules to rearrange and combine the words taken from the prompt

  Incorrect

  Rule-based text processing predates modern language models. LLMs learn relationships between tokens from training data rather than following hand-written rules.

LLMs are trained to generate completions: given a sequence of tokens, the model uses what it learned about how tokens relate to predict the most probable next one. Repeating that step produces the full response.

• AAn image generation model, such as gpt-image-2

  Incorrect

  Image generation models create pictures from prompts. They don't produce vectors for comparing text.

• BA reasoning model, such as gpt-5

  Incorrect

  A reasoning model answers with generated text. Turning each description and query into a comparable vector is the job of a different model type.

• CAn embedding model, such as text-embedding-3-small

  Correct

  Embedding models turn text into vectors that capture meaning, so texts with similar meaning land close together even when they share few words.

• DA speech to text model, such as gpt-4o-mini-transcribe

  Incorrect

  Transcription models convert spoken audio into text. Both the queries and the descriptions here are already text.

An embedding is a vector of numbers that represents the meaning of text. Similar text produces vectors that are close together, which makes embeddings the basis for semantic and vector search.

• AGlobal Standard

  Correct

  Global Standard is the recommended default: pay per token, the lowest price, the broadest region coverage, and new models arrive there first.

• BRegional Provisioned

  Incorrect

  Provisioned deployments reserve throughput that you pay for whether you use it or not, and regional types keep processing in a geography. Neither need applies here.

• CGlobal Batch

  Incorrect

  Batch deployments process large asynchronous jobs at a discount within about 24 hours. A live chat feature needs immediate responses.

• DData Zone Standard

  Incorrect

  Data Zone types keep processing inside a data zone such as the EU or US. With no residency requirement, that restriction adds nothing.

Microsoft's guidance is to start most workloads on Global Standard and move to another deployment type only for a specific reason, such as data residency, reserved throughput or asynchronous batch processing.

• ARaise the top_p parameter

  Incorrect

  top_p is another randomness control. Raising it widens the choice of tokens, which makes answers more varied, not less.

• BMove the deployment to another region

  Incorrect

  The region decides where the model runs, not how consistently it words its answers.

• CSwitch to a Global Batch deployment

  Incorrect

  Batch deployments process requests asynchronously at lower cost. The same model still produces the same variety of answers.

• DLower the temperature parameter

  Correct

  Temperature controls randomness. A lower value makes output more focused and repeatable, while a higher value makes it more varied.

Temperature trades creativity for determinism. Lowering it makes the model choose more probable tokens, so answers become more focused and consistent; Microsoft recommends adjusting temperature or top_p, not both at once.

• AThe AI Red Teaming Agent

  Incorrect

  The AI Red Teaming Agent probes a model or app with simulated attacks to find safety risks. It doesn't compare candidate models on cost or throughput.

• BModel leaderboards in the Foundry portal

  Correct

  Leaderboards compare catalog models on quality, safety, cost and performance benchmarks, and link to each model's detailed results where you can deploy it or try it with your own data.

• CEvaluation runs on the team's own test data

  Incorrect

  Evaluations measure how a model performs on your specific data and use case. That's the step after a shortlist, and the team wants a benchmark comparison before using its own data.

• DThe Foundry agents playground

  Incorrect

  The agents playground tests an agent's instructions, tools and knowledge. It doesn't rank models on cost or throughput.

The model leaderboards compare models from the Foundry catalog using industry benchmarks for quality, safety, cost and performance. Once a model looks suitable, you can open its detailed results, try it in the playground or evaluate it on your own data.

• AA speech to text model for the photo questions

  Incorrect

  Transcription models turn spoken audio into text. They can't read a photo.

• BA vision-enabled chat model for the photo questions

  Correct

  Vision-enabled models such as the GPT-5 series accept images together with text and answer questions about what the image shows.

• CAn image generation model for the photo questions

  Incorrect

  Image generation models create or edit pictures. They don't return a written answer about which panels are dented.

• DA vision-enabled chat model for the newsletter illustrations

  Incorrect

  Vision-enabled chat models read images and respond with text. They don't produce new illustrations.

• EAn image generation model for the newsletter illustrations

  Correct

  Image generation models such as the gpt-image series create new images from text prompts.

Pick the model by the input and output it must handle. Questions about an image need a model that accepts image input and answers in text; creating a picture from a description needs an image generation model.

• AIts token ID from the model's vocabulary

  Incorrect

  Each token keeps the same ID wherever it appears, so "dog" has the same ID in both sentences and the IDs alone can't show order.

• BA positional encoding showing where the token sits

  Correct

  Positional encoding tells the model where each token sits, because order changes how tokens relate to each other.

• CExtra vector dimensions that hold more attributes

  Incorrect

  More dimensions can encode more about a token's meaning, but the same token still gets the same values in both sentences.

• DA sub-word split of the token into smaller parts

  Incorrect

  Sub-word tokens change how text is broken up. They don't record where a token appears in the sequence.

Token vectors are fed into the transformer together with a positional encoding that indicates where each token appears in the sequence, since the order of tokens affects how they relate to one another. Token IDs and the vector values themselves are the same for a token wherever it appears.

• AIt retrains the model's weights each time a policy document changes

  Incorrect

  RAG doesn't change the model's learned parameters; it adds retrieved content to the prompt.

• BIt stores each answer and replays it until someone deletes it

  Incorrect

  Replaying stored answers would keep outdated information; RAG retrieves fresh content for each question.

• CIt raises the model's temperature so answers drift toward new information

  Incorrect

  Temperature affects randomness, not what the model knows.

• DIt retrieves current policy text per request, so updating the index is enough

  Correct

  RAG supplies knowledge at run time without changing the model's parameters, so refreshing the source and index keeps answers current.

RAG retrieves relevant information from selected sources and adds it to the prompt at run time. Because the model's parameters don't change, keeping the documents and index up to date keeps the answers current.

• AThe decoder block

  Incorrect

  The decoder uses embeddings that the encoder has already calculated to produce the model's output. It isn't where the embeddings are created.

• BThe encoder block

  Correct

  In this simplified view, the encoder block creates the embeddings: new vectors with linguistic and semantic characteristics based on the contexts in which each token appears.

• CThe tokenizer

  Incorrect

  The tokenizer breaks text into tokens and assigns each one an integer ID. It doesn't capture meaning from context.

• DThe cosine similarity step

  Incorrect

  Cosine similarity measures how close two existing embeddings are. It compares vectors after they're created and doesn't create them.

Text is first split into tokens with unique IDs, and each token gets an initial vector. The transformer's encoder block transforms these into embeddings that reflect how tokens relate to one another in context. Tokens used in similar contexts end up with similar embeddings, which you can compare with cosine similarity.

• AA model trained from scratch on one company's data for one task

  Incorrect

  A narrow model built only on one company's data is the opposite of a broad, pretrained foundation model.

• BA set of hand-written rules that each agent in a project must follow

  Incorrect

  Rules or instructions shape behavior, but they aren't a model.

• CA large pretrained model with general skills, usable as is or customized

  Correct

  Foundation models such as GPT, Claude and Mistral provide general capabilities out of the box and can be deployed or fine-tuned.

• DA model specialized for producing embeddings for search

  Incorrect

  Embedding models are one model type. A foundation model provides broad, general capabilities.

Foundation models are large, pretrained models that provide general language, reasoning or multimodal capabilities. They serve as the base layer for AI applications and can be deployed immediately or customized.