Part 2 Full Quiz
15 scenario-based questions covering all Part 2 topics. Select an answer to reveal the explanation.
Which architectural approach best meets these requirements?
B is correct. RAG is the correct pattern for enterprise Q&A with traceable source citations. Documents are chunked, embedded, and stored in Azure AI Search. At query time, the most relevant chunks are retrieved and included in the prompt — the LLM answers based on them. You return both the answer and source citations (document name, page). Fine-tuning (A) doesn't make the model "know" specific documents and 800k documents is impractical to fine-tune. Option C is impossible — 800k documents vastly exceeds any context window. Keyword search (D) misses semantic queries and doesn't provide grounded AI answers.
In a RAG pipeline implemented as a Prompt Flow, what is the correct sequence of nodes?
C is correct. In a RAG Prompt Flow the logical sequence is: (1) Search/Index Lookup Node — query Azure AI Search with the user's question to retrieve relevant document chunks; (2) Prompt Template Node — format the retrieved chunks + user question into a grounded prompt using Jinja2 variables; (3) LLM Node — send the formatted prompt to GPT-4o and generate the grounded answer. You retrieve first, format second, generate third. The other options have the sequence wrong — notably option A where the LLM runs before retrieval, meaning it has no context to ground its response.
Which two evaluation metrics are most likely low?
B is correct. Issue 1 — responses contain claims not in source documents = hallucination = low Groundedness. Groundedness specifically measures whether responses are supported by retrieved documents. Issue 2 — "awkward and stilted" language = low Fluency. Fluency measures natural, grammatically correct language. The scenario says responses ARE relevant (ruling out Relevance) and presumably logical (not Coherence). Relevance = does it answer the question (yes). Coherence = is it logically consistent (yes). Fluency = is the language natural (no).
A company wants all three development teams to share the same Azure OpenAI resource connection in Microsoft Foundry, while each team has its own isolated workspace for building their AI solutions. Where should the connection be configured?
B is correct. The Foundry Hub is designed exactly for this — shared infrastructure (connections, storage, compute) is configured once at Hub level and all Projects within that Hub inherit it. Each team gets their own Project (isolated workspace) within the shared Hub. Option A creates separate connections per team — defeating the purpose of sharing. Option C creates unnecessary duplication and cost. Option D (Azure Policy) governs compliance rules, not Foundry connections.
What is the key tradeoff between small and large chunk sizes?
B is correct. The chunk size tradeoff is a core RAG design decision. Small chunks (256 tokens): retrieval is more targeted and precise, but a chunk may lack the surrounding context needed to answer completely. Large chunks (2,048 tokens): more context per chunk, but retrieval is less precise (may include irrelevant information) and multiple large chunks quickly consume the context window. The optimal size (typically 512–1,024 tokens) balances precision and context. Option C is wrong — more information isn't always better; irrelevant content can confuse the model. Options A and D describe fictional constraints.
client.chat.completions.create(model="gpt-4o", messages=[...])The deployment in Foundry is named "contoso-gpt4". The call returns a 404 Not Found error.
What is the cause of the error?
B is correct. In Azure OpenAI, the model= parameter must be the deployment name — the name you assigned when deploying in Foundry or Azure OpenAI Studio. Here that's "contoso-gpt4", not "gpt-4o". Azure looks for a deployment called "gpt-4o", finds none, and returns 404. This is the most common Azure OpenAI mistake. Missing api_version (A) causes a different error. Azure OpenAI uses openai.azure.com not cognitiveservices.azure.com (C). GPT-4o is fully supported in Azure OpenAI (D).
A developer wants to build a multi-turn chatbot that remembers what was said earlier. After the first exchange, the user asks a follow-up question that references the previous answer. What must the developer do to ensure the model understands the context?
B is correct. Azure OpenAI models are stateless — each API call is completely independent with no memory of previous calls. The developer must maintain conversation history in their application code and include the full history (system message + all previous user messages + all previous assistant responses + new user message) in every call. There is no built-in "memory" feature (A). There is no conversation_id parameter (C). max_tokens controls output length, not memory (D).
Which Azure OpenAI capability should be used?
B is correct. The task requires both image understanding AND natural language generation — "describe the product's appearance and suggest use cases." GPT-4o with vision (multimodal) accepts an image and generates a contextual natural language description. DALL-E (A) generates new images — it cannot analyse existing ones. Whisper (C) transcribes audio — there's no audio in this scenario. text-embedding-ada-002 (D) creates vectors for search — it doesn't generate social media text.
A company uses Azure OpenAI to generate marketing images with DALL-E 3. They generate an image, store only its URL, and plan to display it on their website 48 hours later. What will happen?
B is correct. DALL-E image URLs are temporary — they expire after a short period. The correct approach is to download the image bytes immediately after generation and store them permanently in Azure Blob Storage (or similar). Only storing the URL for later use is a common mistake that leads to broken images. This is a tested operational fact about DALL-E integration in Azure OpenAI.
Which Azure OpenAI model should they use for transcription?
B is correct. Whisper is the Azure OpenAI model for audio transcription. It supports 57 languages including Spanish, provides high accuracy, and integrates with Azure OpenAI infrastructure. GPT-4o (A) is multimodal for text+images — it does not process raw audio files. The embedding model (C) converts text to vectors, not audio to text. DALL-E (D) generates images — it has no audio capability.
Which parameter adjustment will most effectively reduce the repetitiveness?
B is correct. frequency_penalty specifically targets word repetition within a response. When set to a positive value (0.5–0.8), the model is penalised for reusing words it has already generated, naturally varying its phrasing. Increasing temperature (A) adds general randomness which may reduce consistency (the chatbot already gives consistent factual answers — breaking that may create new problems). Reducing max_tokens (C) makes responses shorter but doesn't address repetitiveness in the words used. top_p 0.1 (D) makes responses more deterministic — likely making repetition worse.
Which approach most reliably enforces this output format?
B is correct. Fine-tuning is the right choice when you need to permanently bake a specific output format, style, or terminology into the model. Training on JSONL examples that all use Summary/Details/Next Steps headers and formal language teaches the model to always respond this way — without needing to specify the format in every prompt. RAG (A) retrieves data, not formatting rules — using it this way is a misuse. Temperature 0 (C) ensures consistency but doesn't teach a specific format. Stop sequences (D) would stop generation when "Next Steps:" appears — the opposite of what's needed.
Which solution is most appropriate for this scenario?
B is correct. The scenario describes variable traffic — sometimes low, sometimes high. PTU (A) is for predictably high consistent traffic where you need guaranteed throughput — for variable traffic you'd be paying for idle capacity most of the time. The right approach for variable traffic with occasional throttling is: (1) implement exponential backoff retry (handles transient 429s gracefully) and (2) request a higher TPM quota to raise the ceiling. Internet speed (C) has no effect on API quotas. Multi-region (D) splits the traffic but each region has its own quota — it adds complexity without solving the root cause.
Which prompt engineering technique will most effectively improve accuracy?
B is correct. Chain-of-thought (CoT) prompting is the proven technique for improving accuracy on mathematical and multi-step reasoning tasks. By instructing the model to "solve step by step" and "show all calculations and units", it externalises its reasoning — working through mole calculations, unit conversions, and product ratios explicitly rather than jumping to an answer. This dramatically reduces calculation errors. Zero-shot (A) explains rules but doesn't force step-by-step reasoning. Temperature 1.5 (C) increases randomness — catastrophic for maths that needs precision. Fine-tuning (D) might improve chemistry knowledge but won't fix calculation errors caused by skipping steps.
Which combination of approaches correctly addresses ALL four requirements?
B is correct. Breaking down each requirement: (1) Research documents updated weekly → RAG (not fine-tuning — you can't retrain weekly; RAG updates the index without retraining). (2) Deterministic answers → temperature=0 (same prompt → same answer every time). (3) 100k daily requests without throttling → Provisioned Managed (PTU) — guaranteed throughput, no 429s. (4) Reliable section headers every time → fine-tuning on format examples (most reliable method for consistent structure). Option A fails on (1) — fine-tuning can't be updated weekly — and (3) — Standard may throttle at this volume. Option C fails on (2) — temperature 2.0 is highly random. Option D fails on (1) for the same weekly update reason.