Task 2.1: Build generative applications by using Foundry

• APoint the OpenAI SDK at the resource's /openai/v1 route with an API key

  Incorrect

  The API key breaks the no-keys rule, and that route isn't the project endpoint that the agent and evaluation APIs live on.

• BCall the Azure Resource Manager endpoint of the Foundry resource with a token

  Incorrect

  Resource Manager creates and configures resources; it doesn't run agents or return evaluation results.

• CCreate an AIProjectClient on the project endpoint with DefaultAzureCredential

  Correct

  The Foundry SDK exposes every project API, including agents and evaluations, through the single project endpoint and works with Entra ID credentials.

• DUse the Foundry Tools SDK for each service, each with its own service endpoint

  Incorrect

  Foundry Tools SDKs target prebuilt services such as Speech or Vision, each on its own endpoint, not the project's agents.

For agents, evaluations and other project features, use the Foundry SDK with the project endpoint, https://<resource-name>.services.ai.azure.com/api/projects/<project-name>, and a token credential.

• AChange the content filtering configuration of the deployment

  Incorrect

  Filtering settings don't change which APIs the deployed model supports.

• BCall it through the Chat Completions API instead

  Correct

  Not every Foundry model supports the Responses API. The docs say to use Chat Completions when a deployment returns this error.

• CSwitch the app from API key to Microsoft Entra ID authentication

  Incorrect

  The authentication method doesn't add Responses API support to a model that lacks it.

• DRaise the rate limit configured on that model deployment

  Incorrect

  Rate limits govern request volume, not which APIs a model supports.

The Responses API works with Azure OpenAI models and with the Foundry models that support it. When a deployment answers 400 "Model not supported", send its requests through Chat Completions.

• ASwitch the index from hybrid queries to keyword-only queries for precision

  Incorrect

  Retrieval isn't the issue here; changing the query type doesn't stop the model from adding details.

• BInstruct it to answer only from the passages, with citations, or say it can't

  Correct

  The RAG guidance for hallucination despite grounding is to enable citations and use clear instructions to stay with retrieved content.

• CIncrease the passages added to each prompt from 5 to 50 so more facts are present

  Incorrect

  Retrieval is already good; adding passages doesn't stop the model from going beyond them.

• DFine-tune the model on the passages so that it learns the facts in them

  Incorrect

  Training on the passages won't keep each answer inside the text retrieved for it, and it gives no citations.

When retrieval works but answers still drift, tighten the generation step: tell the model to use only the supplied passages, cite them and admit when the answer isn't there. Measure the result with a groundedness check.

• ACoherence

  Incorrect

  Coherence rates logical flow; a coherent answer can still invent content or miss the question.

• BSimilarity

  Incorrect

  Similarity compares the response with a reference answer, which this test set doesn't have.

• CGroundedness

  Correct

  Groundedness checks whether the response stays within the given context without fabricating content, and needs no ground truth.

• DRelevance

  Correct

  Relevance judges how well the response addresses the query, using only the query and response.

• EFluency

  Incorrect

  Fluency rates readability, not whether content is supported or on topic.

Pick evaluators by the data you have and the failure you're hunting. Without reference answers, Groundedness (precision against the context) and Relevance (does it answer the query) cover fabrication and off-topic replies.

• AStructured Outputs with a JSON schema for the order details

  Correct

  Structured Outputs constrain the model to the JSON schema you supply, which JSON mode doesn't guarantee.

• BA higher temperature so the model explores more field names

  Incorrect

  More randomness makes field names less predictable, not more.

• CA larger maximum output token setting

  Incorrect

  Output length isn't the problem; complete JSON with the wrong fields is.

• DFew-shot examples alone, with JSON mode left on

  Incorrect

  Examples help, but JSON mode only guarantees valid JSON, not a schema, so drift can still happen.

JSON mode promises parseable JSON. Structured Outputs go further and make the response follow a JSON Schema, which is what a downstream step that reads named fields needs.

• ALower the reasoning effort to its minimum value for every request

  Incorrect

  Less reasoning shortens requests but weakens answers to hard questions, and nothing guarantees it beats the timeout.

• BSet store to false so that the service doesn't keep the response

  Incorrect

  Background mode requires store=true; stateless requests aren't supported for it.

• CRaise max_output_tokens so that the answer finishes sooner

  Incorrect

  A higher cap allows more output; it doesn't shorten processing or keep a connection open.

• DBackground mode, then poll the response by ID until it completes

  Correct

  Background tasks run asynchronously; you poll for completion or cancel them, so no connection has to stay open.

For long reasoning tasks, start a background response and poll it by ID. Background mode needs stored responses, so keep store set to true.

• AEnvironment variables loaded from one shared configuration file in the repo

  Incorrect

  A shared file is still a copy of the endpoint and key in each deployment, outside the project's governance.

• BA project connection to the search service, read via the project client

  Correct

  Connections link a project to services such as Azure AI Search, so apps get the target from the project instead of local copies.

• CA vector store created by the agent's file search tool for the index

  Incorrect

  A vector store doesn't connect the app to its existing Azure AI Search index.

• DA separate Foundry resource for each environment and its search index

  Incorrect

  More resources multiply the configuration; they don't give one place to manage the link.

Project connections are the Foundry way to wire a project to external resources such as Azure AI Search, with credentials kept by the platform (optionally in your Key Vault). Apps resolve them through the project.

• AIntent Resolution, rating whether the user's goal was met

  Incorrect

  That is an agent evaluator for conversations.

• BFluency alone, rating how natural each translation sounds

  Incorrect

  Fluency uses an LLM judge and ignores the reference translation.

• CGroundedness, comparing each translation with retrieved context

  Incorrect

  Groundedness checks answers against retrieved context, not against a reference translation.

• DA textual similarity evaluator such as BLEU

  Correct

  BLEU measures n-gram overlap between the response and ground truth and is a classic translation quality measure.

Textual similarity evaluators (F1, BLEU, GLEU, ROUGE, METEOR) compare output with ground truth algorithmically; pick them when references exist and a deterministic score is wanted.

• AThe Groundedness evaluator against retrieved context

  Incorrect

  There's no retrieved context here; the check is an exact label.

• BThe string check grader, comparing outputs with labels

  Correct

  string_check performs deterministic comparisons for exact match validation against a reference.

• CThe Relevance evaluator run on each routed message

  Incorrect

  Relevance is an LLM-judged quality score.

• DThe Coherence evaluator with an LLM judge model

  Incorrect

  Coherence judges logical flow, not exact label matches.

Azure OpenAI graders include deterministic graders (string check, text similarity) and model-based ones (labeler, scorer). Use deterministic checks when the right answer is a fixed string.

• AEncode each image as base64 and send it as input

  Correct

  Image input accepts a base64-encoded image as well as image URLs and file IDs.

• BDescribe the screenshot in text and send the description

  Incorrect

  A description loses the visual details the model should see.

• CUpload each image to a vector store for the file search tool

  Incorrect

  Vector stores index documents for retrieval, not image input.

• DSend only the file path of each image on the user's disk

  Incorrect

  The service can't read a local path on the user's machine.

Multimodal requests can carry images by URL, base64 data or an uploaded file ID; pick the one that fits where the image lives.

• Aazure-ai-projects 2.x, the stable line for Foundry (new)

  Correct

  The 2.x SDK targets Foundry (new); 1.x is the stable line for Foundry (classic).

• BThe azure-ai-inference package instead of azure-ai-projects

  Incorrect

  azure-ai-inference retired on August 26, 2026.

• CAny version, since every release supports both portals

  Incorrect

  The docs map SDK versions to portal versions.

• Dazure-ai-projects 1.0.0, as pinned in the sample

  Incorrect

  1.0.0 targets Foundry (classic) projects.

Match the SDK to the platform: azure-ai-projects 2.x for Foundry projects in the new portal, 1.x for classic.

• ATurn off tracing so that the response is no longer kept

  Incorrect

  Tracing is separate telemetry; it doesn't delete stored responses.

• BDelete each stored response by ID through the API

  Correct

  Response data is retained for 30 days by default, and a stored response can be deleted by ID.

• CRedeploy the model so that its stored data resets now

  Incorrect

  Deployments don't hold stored responses.

• DWait 30 days for the data to expire on its own

  Incorrect

  That ignores the user's request to remove it now.

Stored responses can be retrieved, listed and deleted by ID. Plan data removal requests alongside your own telemetry and storage.