Microsoft AI-103 Foundry Model And Service Selection Practice Test

 

Topic 01 covers foundry model and service selection for the Microsoft Certified: Azure AI Apps and Agents Developer Associate certification. These original practice questions apply the verified AI-103 objectives to practical decisions and troubleshooting. Select one answer unless a fixed number is requested. Each option includes an explanation of the relevant behavior and scenario constraints. For broader exam preparation, review the Microsoft AI-103 Exam Dumps page.

Question 1

A support-routing prototype has an acceptance gate of at least 96% classification accuracy and a measured 95th-percentile response time no greater than 600 ms. On the same representative test set, small model S scores 97% at 420 ms; large model L scores 99% at 1,400 ms. Both are approved for the data and deployment region. Which selection is justified for the next pilot?

1. Model L, because its higher accuracy compensates for the response time.
2. Select whichever model has more parameters and repeat the same accuracy test.
3. Run S and L sequentially for every request and use L’s final classification.
4. Model S, with continued monitoring against the same acceptance criteria.
5. Model L with streamed output, treating its first token as the completed classification.

Correct Answer: D

Correct Answer

Answer D is correct because s satisfies both measured gates. L improves accuracy but misses the response deadline; the pilot should validate that S maintains its results under real workload conditions.

Incorrect Answers

Answer A is incorrect because the deadline is a separate required gate, not a metric that can be traded away for two additional accuracy points. L has not demonstrated compliance with that gate.

Answer B is incorrect because parameter count does not override the observed latency constraint. The current task-specific results already distinguish the two candidates for this pilot.

Answer C is incorrect because a mandatory second model call adds work and leaves the final decision dependent on L. The supplied measurements do not support this design within 600 ms.

Answer E is incorrect because the measured requirement concerns the completed routing decision. Streaming an early token does not establish that a usable classification arrives within the deadline.

 

Question 2

An assistant must explain which valves isolate a failed pipe segment from an engineering diagram. OCR extracts every label correctly, but it loses the line connections and arrow directions that determine the answer. Which two design choices preserve the evidence needed by the model? Choose TWO.

1. Select a model whose supported inputs include images and validate its diagram interpretation.
2. Replace the diagram with a caption listing its component names but omitting geometry.
3. Provide image content at a tested resolution that preserves the connections and relevant labels.
4. Retrieve a visually similar diagram and answer from that substitute without checking the original connections.
5. Send only the OCR label list to a larger text-only model.

Correct Answers: A, C

Correct Answers

Answer A is correct because the decisive information is visual topology, not missing text. Image-capable inference can examine those relationships, and task-specific validation is needed before relying on its interpretation.

Answer C is correct because the model needs the actual geometry and legible identifiers together. Adequate input quality preserves the relationships that OCR alone discarded.

Incorrect Answers

Answer B is incorrect because this caption preserves the same labels already recovered by OCR while losing the decisive spatial relationships. It does not address the demonstrated failure.

Answer D is incorrect because visual similarity does not establish identical pipe connectivity. The required answer depends on this diagram’s connections, so a retrieved substitute cannot be treated as equivalent evidence.

Answer E is incorrect because a larger text model still lacks the diagram’s connections. It could infer a plausible layout, but that inference would not be grounded in the supplied engineering evidence.

 

Question 3

A team compares models for transforming deprecated Python calls into a new library’s calls. Candidate C is specialized for code; candidate G is a general language model. Both return well-formed patches, but neither has yet been tested on the team’s repository. Which evidence should control the selection for a guarded migration pilot?

1. Compilation success alone, using whichever model produces fewer compiler errors.
2. The percentage of generated identifiers that match the names in the reference patch.
3. Repository-specific compilation and behavioral tests, including cases where an old call must remain unchanged.
4. The fluency of each model’s explanation of why the new library is preferable.
5. Candidate C’s code-specialized label, without running the repository tests.

Correct Answer: C

Correct Answer

Answer C is correct because a syntactically valid patch can still change behavior or rewrite an inapplicable call. Tests tied to the repository distinguish successful transformation from attractive-looking code and validate the specialized model’s claimed fit.

Incorrect Answers

Answer A is incorrect because compilation is useful but cannot detect every semantic change or unnecessary rewrite. The unchanged-call and behavior cases are essential parts of the stated migration task.

Answer B is incorrect because similar names can coexist with incorrect control flow, arguments or behavior. This comparison does not establish that the transformed application still satisfies its requirements.

Answer D is incorrect because a persuasive explanation evaluates prose, not the correctness of the migration. The task requires executable changes that preserve specified behavior.

Answer E is incorrect because specialization is a reasonable reason to shortlist C, but it does not establish performance on this library and codebase. The pilot decision requires task evidence.

 

Question 4

An intake system reads scanned order forms. Its downstream validator requires schema-aligned fields and links back to the document regions supporting extracted values. The team already has a validated Content Understanding document analyzer that supplies these outputs. Which component should produce the extraction result before a separate model drafts a customer-facing summary?

1. An OCR-only text export, discarding layout after recognizing the characters.
2. A document classifier that identifies the order-form category but returns no field values.
3. The validated Content Understanding analyzer, retaining its field and source information.
4. The summary model’s final paragraph, followed by a parser that guesses the field boundaries.
5. A general chat prompt that asks the model to output fields and estimate page coordinates.

Correct Answer: C

Correct Answer

Answer C is correct because the analyzer supplies the required structured extraction and document grounding. Keeping those outputs separate lets the validator trace values before a narrative model uses them.

Incorrect Answers

Answer A is incorrect because recognized text is useful input, but discarding layout removes the source-region information required downstream. OCR output alone is not the stated extraction contract.

Answer B is incorrect because classification can route the form to the right processing path, but it does not supply the schema fields or their supporting regions. The downstream validator still needs extraction output.

Answer D is incorrect because this reverses the required evidence flow. Values inferred from a generated paragraph do not preserve the analyzer’s traceable relationship to the original form.

Answer E is incorrect because prompting for coordinates does not establish reliable source grounding. The team already has a validated extraction component whose output contract meets the validator’s requirements.

 

Question 5

For a single inference request, a benchmarked model family counts system instructions, input and reserved output against one context limit. A workload requires 114,000 input tokens, 8,000 instruction tokens and up to 12,000 output tokens. Truncation and splitting the analysis are prohibited by this test’s requirements. All listed variants otherwise meet the quality gate. What is the smallest listed context capacity that can accommodate the request?

1. 256,000 tokens.
2. 120,000 tokens.
3. 160,000 tokens.
4. 128,000 tokens.
5. 132,000 tokens.

Correct Answer: C

Correct Answer

Answer C is correct because the required capacity is 114,000 + 8,000 + 12,000 = 134,000 tokens. A 160,000-token limit is the smallest offered value above that total; the scenario supplies the accounting convention.

Incorrect Answers

Answer A is incorrect because this capacity would fit, but it is not the smallest listed capacity that meets the requirement. The question asks for the minimum adequate option.

Answer B is incorrect because this does not even hold the full input plus instructions, before any output allowance is included. Counting only the main document would miss part of the request.

Answer D is incorrect because this accommodates the 122,000 input-and-instruction tokens but not the full 12,000-token output allowance. The stated shared limit includes reserved output.

Answer E is incorrect because this is still 2,000 tokens below the required total. A close capacity does not satisfy a workload that explicitly prohibits truncation.

 

Question 6

A multilingual help-desk pilot must achieve at least 90% correct routing in each supported language. English makes up 95% of the test traffic. Model A scores 98% for English and 74% for the other supported language; model B scores 94% and 92%, respectively. Both meet the latency and deployment requirements. Which conclusion follows from the acceptance rule?

1. Model B passes this language-quality gate; model A does not.
2. Model A passes because its traffic-weighted accuracy is higher.
3. Choose A and translate every non-English request without evaluating the changed pipeline.
4. Both pass because their English results exceed the threshold.
5. Neither passes because the two languages have different accuracy values.

Correct Answer: A

Correct Answer

Answer A is correct because b exceeds the threshold in both languages. A’s strong performance in the larger group cannot compensate for a failing language because the gate applies separately to each group.

Incorrect Answers

Answer B is incorrect because the weighting favors English, but the acceptance rule is not an overall average. Model A’s 74% result fails the independently required threshold.

Answer C is incorrect because translation creates a different system whose routing quality has not been measured here. It may be investigated later, but it does not make A pass the current gate.

Answer D is incorrect because english is only one of the supported languages. Ignoring the smaller group changes the stated requirement and hides a measured failure.

Answer E is incorrect because the rule requires a minimum in each language, not identical scores. B satisfies the minimum even though its results differ slightly.

 

Question 7

An organization’s deployment policy requires inference processing to remain within its approved geography. A team shortlists a model in the Foundry catalog and sees that it is offered from a resource in that geography. Before accepting the candidate, which two checks are still necessary? Choose TWO.

1. Treat the region filter in the catalog as proof that every deployment SKU has the same boundary.
2. Verify the exact model version and deployment option available in the approved region.
3. Verify the chosen deployment option’s documented processing boundary against the policy.
4. Approve any version sharing the model family name once one version is available locally.
5. Use a lower output-token limit as proof that requests cannot leave the geography.

Correct Answers: B, C

Correct Answers

Answer B is correct because catalog presence alone does not establish that the required version and deployment option are available for the intended resource. Availability must match the actual selection.

Answer C is correct because the resource location alone is insufficient evidence of every deployment option’s processing behavior. The selected option must satisfy the organization’s explicitly stated boundary.

Incorrect Answers

Answer A is incorrect because the filter helps identify candidates but does not replace the deployment-specific contract. Different options can have different operational and processing properties.

Answer D is incorrect because versions and deployment options must be checked explicitly. A family-level match does not establish that the intended version has the required availability or processing behavior.

Answer E is incorrect because a response-size setting controls output length, not the geographic scope of inference processing. It cannot establish compliance with this policy.

 

Question 8

A public benchmark ranks model P above model Q. In a held-out test of the organization’s refund rules, Q correctly handles restrictive exceptions that P repeatedly overlooks. Both pass the latency and safety gates. The pilot’s primary quality metric is correct application of those rules, including the exceptions. How should the team use this evidence?

1. Prefer Q for this pilot and retain the held-out cases as a regression set.
2. Remove the restrictive cases so that the local ranking reflects the public ranking.
3. Combine P’s public score with Q’s local accuracy as if they shared one scale.
4. Prefer P because a public ranking must outweigh a smaller task-specific evaluation.
5. Select P after showing it the answers to the held-out cases, without creating a new test set.

Correct Answer: A

Correct Answer

Answer A is correct because the organization’s decision metric directly measures the intended task. Q’s demonstrated handling of the decisive exceptions is stronger evidence for this pilot than an aggregate benchmark for other tasks.

Incorrect Answers

Answer B is incorrect because the exceptions are explicitly part of the acceptance objective. Removing them changes the test population and conceals the failure the evaluation was meant to detect.

Answer C is incorrect because the scores describe different tasks and measurement procedures. Combining their raw values would not produce a meaningful comparison against the stated quality metric.

Answer D is incorrect because a broad ranking does not establish superiority on this policy task. The supplied held-out results measure the requirement that controls the pilot decision.

Answer E is incorrect because using the evaluation answers to tune the system contaminates the evidence. A revised P could be evaluated on new held-out cases, but it has not yet demonstrated the required improvement.

 

Question 9

A web form submits one paragraph and asks for a shorter version. The paragraph and style rules are supplied in every request. The application needs no tools, conversation history or agent lifecycle, and the team wants the fewest additional persisted resources. Which Foundry integration is the best starting point?

1. Create a hosted agent with a container dedicated to each submitted paragraph.
2. Build a search index containing every submitted paragraph before generating the shorter text.
3. Create a persistent prompt agent and a separate conversation for every paragraph.
4. Create a toolbox endpoint with a custom rewriting tool that only forwards the same model request.
5. Call a suitable model directly from the existing application.

Correct Answer: E

Correct Answer

Answer E is correct because the request already contains the information needed for one generation step. A direct model call meets the task without creating a persistent agent resource or a tool orchestration layer.

Incorrect Answers

Answer A is incorrect because hosted agents are useful when custom runtime code or orchestration is required. This task adds neither requirement, so the container and agent lifecycle introduce unnecessary moving parts.

Answer B is incorrect because the paragraph is already included in the request, and there is no cross-document retrieval requirement. Indexing it does not solve a missing information problem.

Answer C is incorrect because a prompt agent can generate text, but its persisted configuration and conversation lifecycle provide no required benefit here. The stated preference is to avoid additional persisted resources.

Answer D is incorrect because the wrapper would add a tool integration without adding a needed capability. A tool layer is not required merely to submit a self-contained generation request.

 

Question 10

An internal assistant must answer from approved procedures that change every day. A current, access-controlled Azure AI Search index already contains the procedures. The model sometimes answers from older knowledge instead. Which two additions are needed in the application request path to ground each answer in the available current evidence? Choose TWO.

1. Retrieve relevant procedure passages using the requester’s authorized scope.
2. Fine-tune the model once on last month’s procedures and remove runtime retrieval.
3. Use the nearest employee’s previous answer as the authoritative procedure source.
4. Provide the retrieved passages and their source identifiers to the generation step.
5. Increase the response token allowance while leaving the model request otherwise unchanged.

Correct Answers: A, D

Correct Answers

Answer A is correct because the index can supply current facts, but the application must actually retrieve the passages relevant to this request and respect the caller’s access boundary.

Answer D is correct because retrieval alone does not change a separate model call. The selected evidence must become input to answer generation so the response can be grounded and traced to its sources.

Incorrect Answers

Answer B is incorrect because a one-time training snapshot cannot supply subsequent daily changes. It also does not replace request-specific access controls on the current procedure corpus.

Answer C is incorrect because a prior answer can be stale and may have been produced under different permissions. It is not a substitute for the approved, current procedure evidence.

Answer E is incorrect because a longer output budget does not introduce current procedure facts into the request. The failure is missing grounding evidence, not insufficient space for an answer.

 

Question 11

Maintenance staff describe a fault as ‘the drive keeps hunting,’ while the manuals discuss ‘unstable rotational speed.’ A tested keyword-only candidate stage returns no useful passage for this query. The team needs a service composition that can retrieve conceptually similar passages even without the same words. Which choice directly addresses that gap?

1. Apply semantic reranking only to the empty candidate set from the unchanged keyword query.
2. Apply a scoring profile to boost manual titles within the unchanged keyword results.
3. Use a compatible embedding model and Azure AI Search vector retrieval to supply passages to the answer model.
4. Increase the number of empty keyword result slots returned to the application.
5. Keep the same retrieval stage and fine-tune only the answer model on examples of well-written explanations.

Correct Answer: C

Correct Answer

Answer C is correct because embeddings support similarity beyond identical wording, and vector retrieval can produce candidates missed by the tested keyword stage. The answer model still needs the retrieved evidence to explain the fault.

Incorrect Answers

Answer A is incorrect because a reranker can reorder retrieved candidates but cannot make an empty supplied set contain the missing passage. The retrieval stage needs an additional way to find it.

Answer B is incorrect because a scoring profile can influence the order of matching documents. It does not supply the absent concept-matching retrieval signal when the tested query finds no useful passage.

Answer D is incorrect because a larger result limit cannot create matches when the tested lexical query retrieves none. It changes the allowance, not the retrieval signal.

Answer E is incorrect because improving answer style does not supply the manual passage missed by retrieval. The source-selection stage must gain a signal that connects the different terms.

 

Question 12

A support assistant must keep a conversation across turns, call approved order tools and let Foundry manage the agent runtime. The team can express its behavior through instructions and tool configuration; it has no custom orchestration code to bring. Which agent type best fits these requirements?

1. A hosted agent solely to obtain tool access, despite having no custom code.
2. A prompt agent that stores conversation history but has no order tools configured.
3. A stateless model endpoint and a plan to implement all conversation and tool handling later.
4. A Foundry prompt agent with the approved tools and managed conversation handling.
5. A direct model call that describes order functions but has no component to execute requested calls.

Correct Answer: D

Correct Answer

Answer D is correct because prompt agents are defined through configuration and use the managed agent runtime. That matches the required tools and ongoing conversation without creating custom runtime code.

Incorrect Answers

Answer A is incorrect because hosted agents also use tools, but custom-code hosting is not needed for the stated behavior. Tool access alone does not require choosing the code-based agent type.

Answer B is incorrect because this supplies part of the requirement but leaves the agent without the approved transactional capabilities. Conversation handling alone cannot perform the required order actions.

Answer C is incorrect because that could become a custom application, but it postpones the capabilities the team explicitly wants Foundry to manage. The prompt-agent configuration already fits the given control boundary.

Answer E is incorrect because describing a callable function does not execute the order operation. The design still lacks the managed tool orchestration required by the scenario.

 

Question 13

A knowledge assistant must answer questions across scanned inspection sheets, recorded interviews and training videos. The raw files have not yet been prepared for retrieval. A design review requires a searchable evidence layer and a generated response that cites the selected evidence. Which three capabilities form the required pipeline? Choose THREE.

1. Generate an answer using the retrieved evidence and retain its citations.
2. Use supported modality-aware extraction to turn the source files into usable content and metadata.
3. Retrain the answer model on every source update before allowing any retrieval request.
4. Index only source URLs and creation dates while omitting the extracted observations.
5. Make previous conversation summaries the authoritative corpus for questions about the original files.
6. Index the prepared content with source identifiers and retrieve passages relevant to each question.

Correct Answers: B, F, A

Correct Answers

Answer B is correct because the raw inputs contain information in different forms. Extraction prepares their content for downstream search while preserving the source context needed by later stages.

Answer F is correct because the searchable evidence layer connects a question to a manageable set of relevant source content. Source identifiers allow the final response to point back to that evidence.

Answer A is correct because the response requirement includes synthesis as well as retrieval. Passing the selected content to generation and preserving references links the answer to the evidence used.

Incorrect Answers

Answer C is incorrect because the required searchable evidence path does not depend on changing model weights for every update. This adds a training lifecycle without replacing extraction, retrieval or grounded generation.

Answer D is incorrect because these records can locate files but do not support content-level retrieval over inspection findings, statements or video events. The evidence layer needs the prepared content, not just a catalog of locations.

Answer E is incorrect because conversation summaries can omit or alter evidence and depend on what earlier users asked. They do not replace a searchable representation of the original source material.

 

Question 14

A document analyzer extracts an order quantity of 480. A later summarization model writes 840. The pipeline overwrites its only quantity field with the summary model’s value, so an investigator cannot determine which stage introduced the discrepancy. Which architecture change most directly preserves the evidence needed to diagnose this failure?

1. Keep the source file and rerun extraction during an investigation instead of retaining the original extraction result.
2. Retain versioned extraction records with source references separately from generated narrative outputs.
3. Repeat summarization three times and retain whichever quantity occurs most often.
4. Increase the summarization model size while continuing to overwrite the extraction record.
5. Store only the final summary and the name of the model that generated it.

Correct Answer: B

Correct Answer

Answer B is correct because separate records preserve what the analyzer produced and what the generation step changed. The investigator can trace the discrepancy to a stage instead of losing the original extraction when narrative output is stored.

Incorrect Answers

Answer A is incorrect because a later run does not preserve what the analyzer actually returned in the disputed transaction. The investigator needs the historical stage output, not a reconstruction that may use different processing conditions.

Answer C is incorrect because agreement among generated outputs is not a record of the original extraction. Majority voting can repeat the same mistake while leaving the evidence-loss problem unresolved.

Answer D is incorrect because a different model might reduce errors, but overwriting still destroys the evidence needed to locate a discrepancy. The stated diagnosis requirement concerns data lineage between stages.

Answer E is incorrect because the model name does not recover the analyzer’s original value. Without stage outputs and their source context, the investigator cannot distinguish extraction from generation errors.

 

Question 15

A team already has an agent implemented in its own Python framework. The agent includes a custom scheduling algorithm and an HTTP contract that must be preserved. It wants Foundry to provide the managed endpoint, scaling and agent identity while retaining control of its runtime code. Which approach fits that boundary?

1. Replace the scheduling algorithm with a prompt-agent instruction and assume equivalent execution.
2. Use a toolbox definition alone as the host for the complete Python process.
3. Deploy only the language model and treat its inference endpoint as the complete application endpoint.
4. Use a prompt agent calling the existing service while the team continues hosting and scaling that Python service.
5. Package the existing agent as a Foundry hosted agent using a supported protocol.

Correct Answer: E

Correct Answer

Answer E is correct because hosted agents let the team bring its code and orchestration while the platform supplies hosting capabilities. Selecting the compatible protocol preserves the required application contract.

Incorrect Answers

Answer A is incorrect because instructions do not establish that the custom deterministic scheduling behavior and HTTP contract will be preserved. The team explicitly needs control of the existing runtime code.

Answer B is incorrect because a toolbox groups callable capabilities and their integration. It is not a substitute for the runtime that executes the team’s complete custom agent application.

Answer C is incorrect because model inference does not include the team’s scheduling algorithm or HTTP handling. The custom application code still needs an execution environment.

Answer D is incorrect because this can preserve the code, but it leaves the application’s runtime hosting and scaling with the team. The requirement is to place those runtime responsibilities under Foundry management.

 

Question 16

A maintenance assistant already has a validated text-only model and document-retrieval workflow. The next release must accept spoken questions and speak the answers while retaining that model and its tested text grounding path. Which two additions fit this requirement? Choose TWO.

1. Replace the validated text workflow with a different end-to-end speech model.
2. Relay audio over a WebSocket to the unchanged text-only model endpoint.
3. Add speech-to-text before the existing text request pipeline.
4. Add text-to-speech after the answer text has been generated and checked.
5. Add spoken-language identification without transcribing the spoken question.

Correct Answers: C, D

Correct Answers

Answer C is correct because transcription turns the spoken question into input the validated text workflow can consume. The retrieval and grounding path can remain the one already tested.

Answer D is correct because speech synthesis provides the requested spoken output from the existing answer text. Keeping it after generation preserves the tested content path before audio presentation.

Incorrect Answers

Answer A is incorrect because an end-to-end speech design may be useful elsewhere, but replacing the model and its reasoning path violates this release’s explicit retention requirement.

Answer B is incorrect because changing transport does not change the model’s accepted input modality. The audio still needs conversion into the text expected by the validated pipeline.

Answer E is incorrect because identifying the language does not provide the actual question text. The existing text workflow still cannot retrieve evidence or answer from a language label alone.

 

Question 17

A billing platform exposes a proprietary API that returns JSON records. There is no supported Azure AI Search indexer for this API. The records must be loaded automatically into an existing search index, and the organization’s policy prohibits an additional persisted staging copy. Which ingestion design fits these constraints?

1. Have application code read the API, map the records to the index schema and push document updates.
2. Copy the API records into a new Blob Storage container and use a blob indexer.
3. Mirror the records into a new SQL database and attach a SQL indexer.
4. Configure a built-in indexer to treat the proprietary API as a supported source without an adapter.
5. Ask an operator to export and upload the records manually whenever updates are noticed.

Correct Answer: A

Correct Answer

Answer A is correct because the push model accepts schema-compatible JSON regardless of the original source. Application code can use the proprietary API directly without creating a persisted intermediate store.

Incorrect Answers

Answer B is incorrect because this is a possible ingestion pattern when staging is allowed, but it creates the extra persisted copy that this organization’s policy prohibits.

Answer C is incorrect because the supported SQL source could be indexed, but the mirror is another persisted staging copy. It does not satisfy the stated data-placement constraint.

Answer D is incorrect because the scenario explicitly establishes that this source has no supported indexer. An arbitrary API cannot gain a compatible connector merely by being supplied as an indexer address.

Answer E is incorrect because manual export does not meet the requirement for automatic ingestion. It also relies on an operator detecting changes instead of a defined update process.

 

Question 18

A new one-to-one blob indexer will load policy files into Azure AI Search. Keys use metadata_storage_path, blob versioning is disabled, and only a small subset of files changes each day. The design must process changes incrementally and remove indexed documents when their source files are deleted. Which two design choices are appropriate? Choose TWO.

1. Reset the indexer each night and rely on the reset alone to remove all earlier deletions.
2. Use the blob indexer’s change tracking with stable document keys for updated files.
3. Enable blob soft delete and configure native deletion detection before the first run, with sufficient retention for the indexer to observe it.
4. Assume hard-deleting a blob automatically deletes its search document without a deletion policy.
5. Generate a new index key every time a policy file changes.

Correct Answers: B, C

Correct Answers

Answer B is correct because the indexer can detect changed blobs through their update timestamps. Stable keys preserve the relationship between a source file and its indexed document instead of creating a new identity on every update.

Answer C is correct because source deletion needs an explicit detection strategy. The soft-deleted state must remain visible long enough for the indexer to remove the corresponding search document, and the policy must exist from the start.

Incorrect Answers

Answer A is incorrect because resetting tracking is not a substitute for a configured deletion strategy. Files that disappeared without being observed as deleted can remain represented in the index.

Answer D is incorrect because ordinary change tracking does not by itself track objects that have disappeared. This assumption can leave orphaned search documents that continue to be retrieved.

Answer E is incorrect because changing the key breaks the stable source-to-document relationship. It can create additional indexed documents instead of updating the existing policy record.

 

Question 19

A parts assistant receives requests containing exact identifiers such as ZX-410-B. The search index already has an identifier-preserving searchable field. Tests show that vector-only retrieval sometimes returns a similar part with a different identifier. For an explicit part-number lookup, which retrieval choice most directly preserves the requested identifier?

1. Use the nearest vector result regardless of whether its identifier matches.
2. Semantically rerank part descriptions without validating the identifier of the selected result.
3. Search only the generic product description field and omit the identifier field.
4. Use an exact lexical match on the prepared identifier field before using the returned part data.
5. Lower the vector similarity threshold so more alternative parts are eligible.

Correct Answer: D

Correct Answer

Answer D is correct because the identifier is a precise lookup constraint. Matching it through the field configured to preserve the value avoids treating a semantically similar but different part as the requested object.

Incorrect Answers

Answer A is incorrect because similarity can retrieve related parts but does not establish identity. The observed failure occurs precisely because a nearby vector can belong to a different part number.

Answer B is incorrect because a description can closely resemble another variant’s description. Reranking by meaning does not establish that the selected record is the exact requested part.

Answer C is incorrect because the description may be shared by several variants. Excluding the prepared identifier field discards the exact distinction that controls this lookup.

Answer E is incorrect because a wider similarity range does not enforce the requested identifier. It can increase the set of related but incorrect parts without solving the identity requirement.

 

Question 20

Users search a repair corpus with both part numbers and informal symptom descriptions. Evaluation shows that keyword search preserves the part numbers but misses paraphrased symptoms; vector search finds the symptoms but sometimes loses the exact part-number match. Which retrieval design combines the two observed strengths in Azure AI Search?

1. Require every informal symptom word to match in a keyword-only query.
2. Keep vector-only search and increase the number of returned neighbors.
3. Issue a hybrid query containing full-text and vector components and combine their ranked results.
4. Apply a larger identifier boost to keyword-only results and ignore the symptom evaluation.
5. Run semantic reranking on keyword results while removing the vector component entirely.

Correct Answer: C

Correct Answer

Answer C is correct because hybrid retrieval uses both lexical and semantic signals in one request. This directly addresses the complementary gaps demonstrated by the two single-method evaluations.

Incorrect Answers

Answer A is incorrect because this makes lexical matching more restrictive even though paraphrases were already being missed. It does not add the semantic retrieval capability the test showed was useful.

Answer B is incorrect because returning more neighbors can improve recall but does not introduce a lexical identifier signal. It does not directly combine the two strengths identified by the evaluation.

Answer D is incorrect because the boost can emphasize part-number matches but leaves the demonstrated paraphrase gap unresolved. The requirement is to combine the observed signals, not optimize only one of them.

Answer E is incorrect because reranking can improve the order of available text candidates, but it does not combine the separately tested vector candidate set with the lexical results. Missing semantic candidates remain a retrieval concern.

 

Question 21

A long troubleshooting manual exceeds the embedding model’s input limit. Fixed nonoverlapping chunks sometimes separate an instruction from the exception immediately following it. Retrieved passages also lose their section heading, making similar procedures hard to distinguish. Which two changes should the ingestion design test? Choose TWO.

1. Remove repeated headings to reduce embedding noise while retaining the same nonoverlapping boundaries.
2. Embed the complete manual in one request despite the stated model limit.
3. Respect meaningful section or paragraph boundaries and use measured overlap where a condition would otherwise be split.
4. Increase the answer model’s context window while passing exactly the same retrieved chunks.
5. Retain the parent heading and source-section identity with each retrievable chunk.

Correct Answers: C, E

Correct Answers

Answer C is correct because the failed relationship spans a chunk boundary. Structure-aware boundaries and targeted overlap can keep the instruction and its qualifying evidence available together without sending the entire manual as one input.

Answer E is correct because the missing heading distinguishes otherwise similar procedures. Carrying that context with a chunk helps the downstream system interpret which procedure the retrieved instruction belongs to.

Incorrect Answers

Answer A is incorrect because this optimization removes the section identity already shown to be useful and leaves the instruction-exception split unchanged. It does not address either observed context failure.

Answer B is incorrect because this exceeds the supplied input constraint and risks rejection or truncation. It is not a valid way to preserve context for this model.

Answer D is incorrect because more available capacity does not restore evidence that the unchanged retrieval payload omits. The ingestion and retrieval design must preserve and supply the relevant context.

 

Question 22

A catalog search assistant must normally expose committed product changes within 30 seconds. Its existing indexer runs every ten minutes. The source can emit change notifications, and the application can push schema-compatible documents to Azure AI Search. Which design should the team test first against the end-to-end freshness target?

1. Process change notifications through a retryable push pipeline and measure when each committed change becomes queryable.
2. Cache each generated answer for ten minutes to reduce repeated index queries.
3. Keep the ten-minute schedule and report the model’s response latency as index freshness.
4. Reset and rebuild the full index at the end of each business day.
5. Increase the number of retrieved search results while keeping ingestion unchanged.

Correct Answer: A

Correct Answer

Answer A is correct because event-driven pushes remove the ten-minute polling delay. Measuring query visibility still matters because submitting an update is not itself proof that the complete freshness target was met.

Incorrect Answers

Answer B is incorrect because this may reduce query work, but it can preserve stale answers for longer than the target. It does not accelerate the source-to-index update path.

Answer C is incorrect because response latency measures how quickly the assistant answers, not how current the indexed content is. It does not remove or measure the polling delay.

Answer D is incorrect because a daily rebuild cannot normally expose each committed change within 30 seconds. Reprocessing unchanged data also does not address the required update cadence.

Answer E is incorrect because a query cannot retrieve a change that has not reached the index. Returning more existing documents does not improve ingestion freshness.

 

Question 23

A shared application serves three tenants. Their contracts explicitly require separate search indexes, but permit the same search service. Clients access search only through the application’s authenticated backend. Which design satisfies that particular isolation requirement?

1. Place all documents in one index and apply a tenant filter to each query.
2. Maintain one index per tenant and have the backend select the allowed index from the authenticated tenant identity.
3. Use one shared index and give each tenant a different query key as the only separation.
4. Create separate indexes but let the client supply any index name without backend authorization.
5. Keep one index and ask the generation prompt to ignore documents belonging to other tenants.

Correct Answer: B

Correct Answer

Answer B is correct because separate indexes satisfy the stated storage boundary. Trusted backend routing also prevents a client from choosing another tenant’s index merely by changing a request parameter.

Incorrect Answers

Answer A is incorrect because this can be a useful logical-isolation pattern under a different policy, but it violates the explicit requirement that these tenants not share an index.

Answer C is incorrect because different credentials do not transform one shared index into three separate indexes. This design does not meet the contract’s required boundary.

Answer D is incorrect because the physical index split alone is insufficient if a client can route a request to another tenant’s index. Selection must be bound to the authenticated tenant.

Answer E is incorrect because a prompt does not satisfy the separate-index requirement or establish an access-control boundary. The requested separation must be enforced before unauthorized evidence reaches generation.

 

Question 24

During a conversation, a user asks the assistant to present figures in euros. Later, the user requests a summary of the company’s current travel reimbursement policy. The policy is maintained in a separate authorized knowledge source. Which two information sources should the application use for the new response? Choose TWO.

1. The user’s own conversation context for the requested presentation preference.
2. The model’s training knowledge as the sole source for the current private policy.
3. The currency preference as an update to the organization’s reimbursement policy.
4. The authorized policy knowledge source for the reimbursement rules.
5. A different employee’s conversation as the source of this user’s presentation settings.

Correct Answers: A, D

Correct Answers

Answer A is correct because the currency preference belongs to this user’s interaction and can be carried forward in that context. It should not become a company-wide policy merely because the user stated it.

Answer D is correct because the policy repository supplies the governing organizational facts. A personal preference can affect presentation without replacing the authoritative policy content.

Incorrect Answers

Answer B is incorrect because training knowledge does not establish the contents of the organization’s current private rules. The designated policy source must supply those facts.

Answer C is incorrect because a user request about presentation does not have the authority to change corporate rules. Combining the two would confuse personal conversational state with shared policy knowledge.

Answer E is incorrect because another employee’s dialogue does not establish the current user’s preference and can cross an inappropriate information boundary. The relevant context is the authenticated user’s own interaction.

 

Question 25

A custom agent currently keeps user preferences only in an in-process dictionary. The product requirement is to restore an authenticated user’s preferences after the process restarts and when the user moves to another device. A shared agent deployment must remain in use. Which integration change directly meets the persistence requirement?

1. Increase the memory assigned to the process while retaining the dictionary as the only store.
2. Store authorized preferences in a durable service keyed by the stable user identity and load them when a session starts.
3. Create a new anonymous session identifier on every request and use it as the only preference key.
4. Add the latest user’s preferences to the shared agent’s global instructions.
5. Write the preferences only to the first device’s local browser storage.

Correct Answer: B

Correct Answer

Answer B is correct because an external durable record can survive the application process and can be retrieved from different devices under the same authenticated identity. The application must control which user’s record is loaded.

Incorrect Answers

Answer A is incorrect because more process memory does not survive a process restart. It also does not create a shared identity-based record that another device can retrieve.

Answer C is incorrect because a fresh identifier does not connect the new request to the returning authenticated user’s saved preferences. It fragments state instead of restoring it.

Answer D is incorrect because global instructions would apply one user’s settings to other users of the same deployment. Persistence must preserve user scope as well as survive process restarts.

Answer E is incorrect because local storage can help that browser but does not satisfy the cross-device requirement by itself. The new device needs an authorized way to retrieve the same durable record.

 

Question 26

An assistant displays stock availability before a customer confirms a reservation. The application already has an approved read-only inventory API that returns the authoritative current quantity and its observation time. Search content is refreshed hourly, while reservations change stock throughout the hour. Which integration should supply the availability displayed for this decision?

1. Reuse the last quantity remembered in the customer’s previous conversation.
2. Use the highest-ranked stock figure from the hourly search index without checking its age.
3. Lower the cache lifetime to ten minutes and treat every cached result as current.
4. Ask a model fine-tuned on historical sales to estimate the current quantity.
5. Invoke the authorized inventory tool at request time and identify the result’s observation time.

Correct Answer: E

Correct Answer

Answer E is correct because the live tool can supply the authoritative snapshot for the current request. Reporting its observation time avoids presenting that snapshot as a guarantee that stock cannot change afterward.

Incorrect Answers

Answer A is incorrect because conversation memory records what was known or said earlier. It is not authoritative evidence of inventory after subsequent reservations.

Answer B is incorrect because the indexed value can precede reservations made since the refresh. Retrieval relevance does not make an old quantity current.

Answer C is incorrect because a shorter cache can reduce staleness but still permits changes during those ten minutes. It does not replace the available request-time authoritative read.

Answer D is incorrect because a forecast from historical patterns is different from reading the current inventory state. The approved API already supplies the authoritative observation needed here.

 

Question 27

An agent’s knowledge source contains two versions of a maintenance rule: one applies through September 30 and the next starts October 1. A user asks which rule governed an incident on September 20. Both documents are relevant to the equipment, and the newer version ranks higher in an unfiltered search. Which knowledge-integration contract best supports the answer?

1. Always return the newest document because newer publication dates override the incident date.
2. Use whichever version occurs most often in the user’s conversation history.
3. Retrieve by the incident’s effective date and return the applicable version with its source reference.
4. Remove the older document so that retrieval cannot return conflicting versions.
5. Return only vector similarity scores and let the answer model infer the missing effective dates.

Correct Answer: C

Correct Answer

Answer C is correct because relevance to the equipment is not enough when validity changes over time. The knowledge tool must preserve and apply version metadata so the answer uses the rule that governed the specified incident date.

Incorrect Answers

Answer A is incorrect because the new rule starts October 1, after the incident. Newness does not make it applicable to a September 20 event under the dates given.

Answer B is incorrect because frequency in a conversation does not establish temporal authority. The effective dates in the knowledge source determine which version applies.

Answer D is incorrect because the older rule is the applicable evidence for this historical incident. Removing it would prevent the system from answering the user’s actual question correctly.

Answer E is incorrect because similarity scores do not encode the documented validity period in a form the answer can safely treat as authoritative. The integration should supply the relevant version metadata and evidence.

 

Question 28

An organization’s memory policy requires personal preferences to expire after 30 days and requires an authenticated deletion request to remove that user’s preferences from all active serving stores within 24 hours. The application uses a shared agent deployment and an application-managed memory service. Which three capabilities must the memory integration provide? Choose THREE.

1. Enforce the 30-day expiry requirement on records and their serving path.
2. Keep the stored preferences retrievable and append a prompt telling the model to forget them.
3. Delete the selected user’s records and invalidate their active serving copies on request.
4. Run a monthly full-store cleanup as the only response to individual deletion requests.
5. Compress old preferences into a common summary without retaining ownership or deletion mappings.
6. Associate memory records and serving copies with an authenticated user identity.

Correct Answers: F, C, A

Correct Answers

Answer F is correct because the system needs a reliable way to locate the requesting user’s records without confusing them with another user’s data. Anonymous or deployment-wide records cannot support targeted deletion safely.

Answer C is correct because the 24-hour request requirement applies to every active source from which the agent can recover the preference. Deleting one record while leaving a serving cache available would be incomplete.

Answer A is correct because automatic expiry addresses the separate age-based rule even when a user never submits a deletion request. Expired information must stop being returned by the memory integration.

Incorrect Answers

Answer B is incorrect because a prompt does not remove information from active serving stores. The requirement applies to the memory system’s behavior, not only to an instruction about how the model should respond.

Answer D is incorrect because a request near the beginning of the month could remain unfulfilled far beyond 24 hours. A periodic cleanup alone does not satisfy the separate request deadline.

Answer E is incorrect because compression does not remove the preferences from serving use and discards the identity mapping needed for targeted deletion. It is not a substitute for lifecycle controls.

 

Question 29

A business agent must query a private scheduling API for a fresh result on each request. The API is reachable only from a specified virtual network and requires a scoped application identity. The team may implement a custom tool executor. Which integration design addresses both requirements?

1. Add a private DNS record while leaving the executor with no route into the virtual network.
2. Grant the project administrator a broad management role while leaving the executor outside the permitted network.
3. Run the tool executor with verified private-network reachability and an identity authorized for the API’s required operation.
4. Describe the API’s schema in the prompt and rely on the model to supply results without an executor.
5. Index the API’s responses once a day and replace all live tool calls with retrieval.

Correct Answer: C

Correct Answer

Answer C is correct because the executor needs both a valid network path and permission to perform the call. Treating connectivity and authorization as separate checks prevents assuming that one automatically supplies the other.

Incorrect Answers

Answer A is incorrect because name resolution alone does not establish network connectivity. The executor must be able to reach the resolved private endpoint as well as authenticate to it.

Answer B is incorrect because an administrator’s management permission does not supply the executor with a private route or the runtime API identity. The actual caller still cannot satisfy the stated requirements.

Answer D is incorrect because a schema tells the model what an operation looks like but does not execute it. The authoritative fresh result requires a component that actually reaches and calls the API.

Answer E is incorrect because a daily knowledge snapshot does not satisfy the requirement for a fresh scheduling result on each request. It changes the information contract instead of implementing the private integration.

 

Question 30

Several employees share one agent deployment. A defect causes every request to load memories using only the deployment ID, so one employee sees another employee’s saved preferences. The application already authenticates each employee and must continue using the shared deployment. Which correction addresses the memory-scope defect?

1. Use an employee name supplied in the request body without checking the authenticated identity.
2. Use only the organization’s tenant ID as the memory key for all employees.
3. Clear all employees’ stored preferences at the end of every request.
4. Tell the model to ignore preferences that appear to belong to another employee after loading them.
5. Resolve and enforce the memory partition from the authenticated employee identity before loading context.

Correct Answer: E

Correct Answer

Answer E is correct because the deployment ID identifies the shared application, not the person whose memory is authorized. Binding retrieval to the authenticated employee separates personal context while retaining the shared agent.

Incorrect Answers

Answer A is incorrect because a caller could choose another employee’s name. A display value supplied by the client is not a trustworthy authorization boundary for loading personal memory.

Answer B is incorrect because employees in the same tenant would still share a memory scope. The defect requires individual user separation, not merely separation between organizations.

Answer C is incorrect because this destroys the intended saved-preference feature instead of correcting ownership. The system should retain authorized personal memory and select the proper scope for each employee.

Answer D is incorrect because this leaves the unauthorized context available to the model and relies on inference to repair an access-control error. Scope must be enforced before personal memory is loaded.