Implement generative AI solutions practice questions

Local deployment

Why it fails: Local deployment is not scalable for production.

Azure Batch

Why it fails: Azure Batch is for batch processing, not real-time.

Azure Functions

Why it fails: Azure Functions may have cold start latency.

Fine-tune the model on a dataset of recent code.

Why it fails: Fine-tuning is resource-intensive and may be overkill.

Set the temperature parameter to 0 to reduce randomness.

Why it fails: Temperature does not control factual accuracy.

Provide a few-shot example of correct code in the prompt.

Why it fails: Few-shot examples may not generalize to all cases.

Fine-tune a model to generate the data locally.

Why it fails: Fine-tuning is for customization, not data generation.

Use the streaming API with multiple concurrent connections.

Why it fails: Streaming is for real-time, not bulk generation.

Deploy a model on Azure Functions and call it in parallel.

Why it fails: Azure Functions adds overhead and complexity.

Azure Blob Storage

Why it fails: Blob Storage is for file storage, not search.

Azure SQL Database

Why it fails: SQL Database lacks native vector search.

Azure Functions

Why it fails: Functions is compute, not a data store.

Configure Prompt Shields in Azure AI Content Safety.

Why it fails: Prompt Shields protect against prompt injection attacks, not automatically filter harmful content.

Use Azure AI Content Safety with a custom category severity threshold.

Why it fails: Custom Content Safety is for custom categories, but the question asks for automatic filtering of harmful content, which is built into the default content filtering.

Use groundedness detection in Azure AI Content Safety.

Why it fails: Groundedness detection checks if the model's output is based on source material, not for harmful content.

Set the max_tokens parameter to a high value to allow longer outputs.

Why it fails: Increasing max_tokens does not affect the security of the generated code.

Fine-tune the model on a dataset containing examples of insecure code.

Why it fails: Fine-tuning on insecure code would teach the model to produce more vulnerabilities.

Add a system message that instructs the model to never generate insecure code.

Why it fails: System messages can provide guidance but do not guarantee compliance, and the model may still generate insecure code.

The citation format [^1] is not supported in Azure OpenAI with data sources.

Why it fails: The [^1] format is the standard citation format for Azure OpenAI with data sources.

The system message overrides the citation rendering.

Why it fails: The system message does not affect citation rendering; it only sets the assistant's behavior.

The 'context' object in the assistant response is not being passed back in the prompt for the next turn.

Why it fails: The context object is part of the API response and is not required in the prompt for citations to appear.

Create resources, then configure failover routing, then enable geo-replication, then set up monitoring, then test failover

Why it fails: This is incorrect because failover routing should be configured after replication is set up, not before. Replication ensures data consistency across regions, which routing depends on.

Create resources, then enable geo-replication, then set up monitoring, then configure failover routing, then test failover

Why it fails: This is incorrect because monitoring should be set up after failover routing is configured, as routing defines the endpoints to monitor. Also, routing is needed to direct traffic before monitoring can be meaningful.

Create resources, then enable geo-replication, then configure failover routing, then test failover, then set up monitoring

Why it fails: This is incorrect because monitoring must be in place before testing. Without monitoring, you cannot evaluate the success of the failover test or detect issues during the test.

Step 1: Create the Azure AI Document Intelligence resource, Step 2: Gather and label documents, Step 3: Train the model, Step 4: Test the model, Step 5: Publish the model

Why it fails: This is incorrect because you should gather and label documents before creating the resource, since you need to know the types of documents you're working with. Creating the resource first is premature.

Step 1: Gather and label documents, Step 2: Train the model, Step 3: Create the Azure AI Document Intelligence resource, Step 4: Test the model, Step 5: Publish the model

Why it fails: This is incorrect because you cannot train the model without first having the Azure AI Document Intelligence resource created to host the training process.

Step 1: Gather and label documents, Step 2: Create the Azure AI Document Intelligence resource, Step 3: Test the model, Step 4: Train the model, Step 5: Publish the model

Why it fails: This is incorrect because you must train the model before testing it; testing without training is not possible.

Direct Line: Embeds a chat control in a web page.

Why it fails: Incorrect — this describes Web Chat, not Direct Line.

Microsoft Teams: Enables communication between your bot and your own client application.

Why it fails: Incorrect — this describes Direct Line, not Microsoft Teams.

Set the temperature parameter to 0 to make the output deterministic.

Why it fails: Deterministic output does not ensure security.

Post-process the generated code using a static code analysis tool.

Why it fails: Post-processing is valuable but not the most effective upfront mitigation.

Fine-tune the model on a dataset of secure code examples.

Why it fails: Fine-tuning is resource-intensive and may not cover all vulnerability types.

Enable customer-managed keys for encryption.

Why it fails: Customer-managed keys do not affect model training data usage.

Use a private endpoint to restrict access to the service.

Why it fails: Private endpoints control network access, not data usage.

Configure data residency to keep data in a specific region.

Why it fails: Data residency does not control model improvement.

temperature

Why it fails: Temperature does not specify language.

system message

Why it fails: System message can set behavior but user message should specify language.

max_tokens

Why it fails: max_tokens does not specify language.

max_tokens

Why it fails: max_tokens caps the response length in tokens; it cannot influence tone, only truncation. It is tempting because it governs output size, and would be the right parameter when a poem must fit a strict length or cost budget rather than convey a particular mood.

top_p

Why it fails: top_p narrows nucleus sampling to a cumulative probability mass, affecting diversity rather than tone. It is tempting because it also steers generation, and would be correct when you need to constrain token selection to a probability threshold for consistent, focused output.

temperature

Why it fails: Temperature controls randomness and sampling diversity, not emotional register; a cheerful tone comes from prompt wording or system message instructions. Temperature is tempting because it shapes output style, and would be correct when you need varied, creative versus deterministic responses from the same prompt.

Whisper

Why it fails: Whisper performs speech-to-text transcription and translation, producing no visual output. It is tempting because it is a genuine Azure OpenAI model, but image generation requires DALL-E; Whisper would be correct for transcribing audio, not synthesising a cat wearing a hat.

GPT-4

Why it fails: GPT-4 handles text and multimodal understanding, not image synthesis. It is tempting because it accepts image inputs and is a flagship Azure OpenAI model, but generating a new picture requires DALL-E; GPT-4 would suit describing or analysing an existing image instead.

Codex

Why it fails: Codex generates and completes source code from natural-language prompts, not pictures. It is tempting because it is a real OpenAI model for programmatic output, but the stem needs visual synthesis, which only DALL-E provides; Codex would be correct for writing a function.

Enable customer-managed keys for encryption.

Why it fails: Customer-managed keys do not affect model training data usage.

Use a private endpoint to restrict access to the service.

Why it fails: Private endpoints control network access, not data usage.

Configure data residency to keep data in a specific region.

Why it fails: Data residency does not control model improvement.

temperature

Why it fails: temperature scales the sampling distribution's randomness; it cannot mandate Python syntax. It is tempting because it is the standard parameter for shaping model behaviour, but it is the correct choice when adjusting creativity or determinism, not for specifying the target language.

max_tokens

Why it fails: max_tokens caps the response length in tokens; it cannot influence which programming language the model emits. It is tempting because it is a genuine generation-control parameter, but it is the correct choice only when limiting output size or cost, not for steering language.

top_p

Why it fails: top_p performs nucleus sampling, restricting the cumulative probability mass of candidate tokens; it does not select a programming language. It is tempting because it is a real sampling control, but it is the correct choice when tuning output diversity, not for enforcing Python.

max_tokens

Why it fails: max_tokens does not specify language.

temperature

Why it fails: Temperature does not specify language.

system message

Why it fails: System message can set behavior but user message should specify language.

Set max_tokens to a small value to limit output length.

Why it fails: max_tokens caps response length only; it does not constrain vocabulary, phrasing, or tone, so outputs remain stylistically inconsistent. It is tempting because shorter replies feel uniform, but that parameter suits cost or latency control, whereas style consistency requires system message instructions or few-shot examples.

Fine-tune the model on a dataset of product descriptions.

Why it fails: Fine-tuning alters model weights for task or domain behaviour, not a guaranteed uniform house style across every generation. It is tempting because it genuinely teaches format and tone from examples, and would be right when the model must adopt a niche vocabulary or output structure that prompting cannot reliably produce.

Increase the temperature to 1.0 for more creativity.

Why it fails: Temperature controls sampling randomness, so 1.0 increases variation and produces less predictable wording, directly undermining consistent style and tone. It is tempting because creativity sounds desirable for marketing copy, but high temperature suits brainstorming or varied drafts, not uniform product descriptions.

Use a single Azure OpenAI deployment and configure content filters at the application level.

Why it fails: Content filters are configured per deployment, not per application.

Configure different system messages for each tenant to enforce content policies.

Why it fails: System messages are not a substitute for content filters.

Pass the tenant ID in the API call and use a custom middleware to apply filters.

Why it fails: Azure OpenAI does not support per-request content filter configuration.