AI Intermediate: Azure AI Foundry & Azure OpenAI Practice Questions

The most common operational mistake is creating a custom content filter policy but never assigning it to the specific Azure OpenAI deployment. The deployment then keeps using the default filter, which would explain why severity 2 violent content passes through. Severity level 2 in Azure AI Content Safety maps to 'medium', not 'low'; low is level 1, so the claim that level 2 is 'low' is incorrect. The 'Violence' category IS configurable in Azure OpenAI content filters. It is one of the four standard harm categories (Hate, Violence, Sexual, Self-Harm) whose thresholds can be adjusted. Content filters apply to both the completions and chat completions endpoints, so there is no endpoint-based limit on where a filter applies.

Groundedness is a built-in evaluation metric in Azure AI Foundry / Prompt Flow. It measures whether a model's response is factually supported by the context documents it was given, which makes it the right metric for catching hallucinations in a RAG pipeline. Fluency measures how grammatical and natural the response sounds, which has nothing to do with factual grounding. Coherence measures how logically consistent and readable the response is, but it doesn't check whether the facts come from the source documents. F1 Score measures token overlap between the generated answer and a reference answer. It is a lexical similarity metric and can't tell whether the facts came from the retrieved context or were hallucinated.

The basic structural difference between the legacy completions API and the chat completions API is the input format. Completions takes a single 'prompt' string. Chat completions takes a 'messages' array of objects, each with a 'role' ('system', 'user', 'assistant') and 'content'. This format carries multi-turn conversation context. The 'engine' vs. 'model' naming is a real legacy distinction, but it is a minor field-name change, not the most significant structural change. Both Azure OpenAI endpoints require the 'api-version' query parameter, so it is not unique to chat completions. Both endpoints use JSON request bodies; no form-data encoding is involved.

The root cause is stale data in the Azure AI Search index. Re-indexing removes documents that have been deleted from the source. A scheduled indexer run keeps the index continuously synchronized with the document store as changes happen. Raising the tokens-per-minute quota is wrong because quota affects throughput, not document freshness. Switching to the o1 model is wrong because a different model does not fix the underlying data staleness. Enabling semantic ranking is wrong because it only reorders results by relevance and does not remove deleted or outdated documents from the index.

Azure AI Content Safety returns severity levels on a 0–6 scale (0, 2, 4, 6). A severity of 0 means no harmful content was found in that category, so the content is considered safe on that dimension. Exceeding the maximum input length is wrong because length problems return errors, not a severity of 0. A blocklist match is wrong because blocklist matches are reported separately from severity scores. A classifier failure is wrong because a severity of 0 is a valid, meaningful result, not an error condition.