Generative AI — Models, Deployment and Prompts — Azure AI Apps and Agents Developer AI-103 practice questions

What is the best way to debug a skillset that writes the wrong value into an index field?

1. Delete and rebuild the index so the enrichment tree is regenerated cleanly
2. Restart the search service from the portal to clear cached skill outputs
3. Run a Debug Session to step through each skill's inputs and outputs
4. Turn off role-based access so the skillset can read every source field

Answer: C — Run a Debug Session to step through each skill's inputs and outputs

C) Right — a Debug Session replays one document through the enrichment pipeline so you can inspect each skill's inputs, outputs and output field mappings. A) Rebuilding the index reindexes the same faulty projection. B) The service cannot be restarted, and stale cache is not the cause. D) Loosening access control hides nothing about mapping errors and weakens security.

The groundedness evaluator in Microsoft Foundry helps a team detect what?

1. Network latency between the app and the model endpoint
2. Answers not supported by the supplied grounding sources
3. Cache hit rates for repeated prompts across a session
4. Token consumption per deployment against its rate limits

Answer: B — Answers not supported by the supplied grounding sources

B) Groundedness compares each claim in the response with the retrieved context and flags unsupported statements, which is the primary hallucination signal for a retrieval-augmented application. A) Latency is an operational metric from tracing, not an evaluator. C) Caching is a cost technique carrying no quality signal. D) Token analytics track spend and throttling, not correctness.

What is a sensible baseline chunking strategy for long PDFs feeding a retrieval index?

1. Split at random character offsets so chunks vary and cover every boundary
2. Embed each document whole so no context is ever lost between the chunks
3. Emit one token per chunk so retrieval can pinpoint the exact match found
4. Use page or few-hundred-token chunks with overlap, embedded individually

Answer: D — Use page or few-hundred-token chunks with overlap, embedded individually

D) Right — moderate chunks with a small overlap keep each vector topically coherent while preserving context across boundaries. A) Random splits cut sentences and blur the embedding. B) A whole-document vector averages away the specific passage you need. C) Single-token chunks carry no usable semantics.

When is fine-tuning typically more expensive than retrieval-augmented generation?

1. Only for image generation workloads, where training runs are very long
2. Never — retrieval always costs more once embedding storage is included
3. The two approaches cost the same, since both bill per processed token
4. At low to medium scale, where retrieval avoids training and hosting cost

Answer: D — At low to medium scale, where retrieval avoids training and hosting cost

D) Right — retrieval adds embedding and index cost but skips training runs and dedicated hosting, so it is usually the cheaper starting point. A) The comparison is not specific to image workloads. B) Reverses the usual economics. C) Fine-tuning adds training and hosting charges on top of token billing.

An Azure Search in Foundry Tools indexer must enrich blobs with OCR and entity recognition. Where is that chain defined?

1. In a Logic App that runs between the blob store and the search index
2. In a SQL trigger that fires whenever a new blob is uploaded
3. In a skillset attached to the indexer that runs the enrichers
4. In a Power BI dataflow that pre-processes each blob on import

Answer: C — In a skillset attached to the indexer that runs the enrichers

C) A skillset lists the cognitive skills the indexer runs during ingestion and maps their output into index fields. A) An orchestration app can move data but does not drive indexer enrichment. B) Blob ingestion is not triggered from a relational database. D) Reporting tools consume an index rather than build one.

Making content filters more permissive on an Azure OpenAI in Foundry Models deployment requires:

1. Toggling a single per-deployment switch in the Foundry portal
2. An approved modified content filter request and eligibility review
3. Purchasing provisioned throughput units for the model deployment
4. Calling the public OpenAI API instead of the Azure endpoint

Answer: B — An approved modified content filter request and eligibility review

B) Relaxed filter settings are gated: you submit a modification request and must meet the eligibility criteria before the looser policy can be applied. A) No self-service toggle loosens the defaults. C) Throughput purchasing is a capacity decision. D) Leaving Azure abandons the governance the question assumes.

For RAG over private documents, which Azure Search in Foundry Tools index design fits best?

1. A keyword-only index with a rich set of scoring profiles
2. A SQL view queried directly from the application tier
3. No index; send the whole corpus in the prompt each turn
4. A vector index of embeddings, queried with hybrid search

Answer: D — A vector index of embeddings, queried with hybrid search

D) Vector fields let retrieval match on meaning rather than exact words, and hybrid search combines that with keyword scoring and semantic ranking for the best recall. A) Keyword alone misses paraphrase and synonyms. B) A SQL view offers no embedding or relevance ranking. C) A corpus exceeds the context window and costs far more per call.

A generated answer states facts absent from the retrieval sources supplied to the model. Which evaluation concept describes this?

1. Poor groundedness — claims unsupported by the retrieved source material
2. Low fluency, since the generated sentences read awkwardly to a reader
3. Excessive latency between the request and the first streamed response
4. Token overflow, where the prompt exceeds the model's context window

Answer: A — Poor groundedness — claims unsupported by the retrieved source material

A) Right — groundedness measures whether each claim is supported by the supplied context, and unsupported claims fail it. B) Fabricated text is often perfectly fluent, so fluency misses the problem. C) Latency is a performance signal, unrelated to truthfulness. D) Context overflow is an input-size error, not invented content.

In an Azure Search in Foundry Tools index, how do the 'searchable' and 'filterable' attributes differ?

1. Neither attribute is needed once the field has been marked retrievable
2. They are aliases, so setting either one enables the same query behaviour
3. Searchable enables analyzed full-text queries, filterable the $filter use
4. Both exist only to enable $orderby sorting over the values in the field

Answer: C — Searchable enables analyzed full-text queries, filterable the $filter use

C) Right — searchable runs the field through an analyzer for full-text matching, while filterable stores it for exact $filter predicates. A) Retrievable only controls whether the value is returned in results. B) The two attributes drive different query paths and are set independently. D) Sorting is governed by the separate sortable attribute.

Does using Azure OpenAI in Foundry Models require separate Microsoft approval?

1. Yes, and image generation models remain blocked for all customers today
2. No, and no feature of the service has ever required a registration form
3. Yes, an approval form is required before any embedding model can be used
4. Access is general for most customers; limited-access features apply

Answer: D — Access is general for most customers; limited-access features apply

D) Right — the general registration requirement was removed, but limited-access capabilities and modified abuse-monitoring or content-filter settings still need a review. A) Image generation is generally available. B) An application process did exist historically. C) Embedding models were never gated behind an approval form.

Define groundedness and relevance as evaluators.

Groundedness measures whether the response is supported by the supplied context, catching fabricated claims. Relevance measures whether the response actually addresses the user's question. A reply can be perfectly grounded in retrieved text and still be irrelevant, which is why both are scored.

Why is a system prompt not a security control, and what compensates?

Instructions and untrusted content share the same context window, so a crafted user message or retrieved document can attempt to override them. Compensating controls are prompt shields, content filters, treating retrieved text as data rather than instructions, output validation, and enforcing authorization in the tool layer instead of in prose.

Deployment types for a model?

Standard, billed per token on shared regional capacity. • Global and data zone variants that route across a wider capacity pool. • Provisioned throughput, reserving dedicated capacity for predictable latency. • Batch, for high-volume asynchronous jobs at lower cost.

Protected Material detection?

Identifies generated content that matches known song lyrics, news, code from training data.