Microsoft AI-103 Practice Test 9
Question 80
A search service retrieves relevant passages but the best evidence appears below several loosely related hits. What post-retrieval step should be evaluated?
- Return search results in the order of their storage-object filenames.
- Rerank the candidate passages for relevance to the actual question before composing context.
- Keep the original vector scores as final ordering without a query-aware reranking stage.
- Generate an answer using only the first retrieved hit regardless of its content.
Explanation
Correct Answer
B:A reranker evaluates the specific query-passage relationship and can improve the quality of the evidence supplied to generation.
Incorrect Answers
A:Filename order does not assess whether the passage answers the user's actual question.
C:Initial retrieval can include loosely related passages and miss the best evidence near the top.
D:The first hit can be less relevant than lower-ranked evidence and needs relevance checking.
Question 81
Support requests use different wording than the troubleshooting documents. What retrieval method helps match similar meaning?
- Use vector similarity retrieval with suitable document embeddings
- Rerank exact-title keyword matches only and omit vector retrieval for differently worded support questions.
- Require exact term equality for every support request and disable vector similarity for paraphrased troubleshooting symptoms.
- Use exact keyword matching only and reject results that lack the user's specific phrasing, even when they convey the same issue.
Explanation
Correct Answer
A:Embeddings enable retrieval by semantic proximity even when question and document use different vocabulary. The practical requirement is to retrieve by semantic meaning rather than exact words; a solution built to rerank relevant candidates for question fit would not meet that same need.
Incorrect Answers
B:A reranker only reorders the candidate set. It cannot retrieve conceptually similar passages missing from keyword results; vector or hybrid retrieval is needed for paraphrases.
C:Exact matching misses semantically similar descriptions that use different words. Vector similarity or hybrid retrieval is better suited to symptom paraphrases.
D:Lexical-only retrieval can miss paraphrases. Embeddings or semantic matching help connect different wording that expresses the same troubleshooting intent.
Question 82
A RAG chatbot must answer from newly updated company policy, cite supporting evidence, and resist hostile instructions embedded in indexed documents. Which THREE runtime practices matter? Choose THREE.
- Wrap retrieved snippets with explicit source identifiers and delimiters in the grounded prompt
- Enable hybrid search alone and skip policy-update reconciliation, citations and retrieved-text injection controls.
- Treat retrieved handbook instructions as higher-priority commands than the application's tool policy, provided the handbook was recently indexed.
- Treat retrieved text as untrusted evidence and block its instructions from controlling the agent
- Keep all deleted and superseded policy passages available to the chatbot without document-version filtering or index change tracking.
- Retrieve current relevant policy passages from an authorized indexed corpus
Explanation
Correct Answers
A:Source separation lets generation and citation checks distinguish evidence from instructions and unrelated passages. Authorized retrieval, source-labeled prompt assembly and untrusted-document isolation jointly protect grounded responses.
D:RAG evidence is lower trust than the application's policy and should never grant extra privileges.
F:RAG grounds generation on up-to-date enterprise content rather than relying on static parametric memory.
Incorrect Answers
B:Hybrid search finds candidate content but does not guarantee policy currency, traceable support or resistance to hostile instructions. Each safety and provenance control needs independent enforcement.
C:A current document is evidence, not an authorized control channel. The RAG system must retrieve current passages, maintain citations and disregard embedded instructions that attempt to redirect the agent.
E:Leaving stale or deleted passages eligible for retrieval can cause obsolete policy answers. Keeping an authorized, current index is separate from the prompt-level injection and citation controls.
Question 83
Technical answers must incorporate both exact error identifiers and descriptive symptom matches. What query design is most suitable?
- Use only a keyword query and discard vector matches for descriptive symptoms written with different terminology.
- Run only vector search after stripping product serials and error codes from queries, preserving generic symptom words.
- Combine lexical and vector queries in hybrid search
- Rely on semantic reranking without searching the exact diagnostic error identifier.
Explanation
Correct Answer
C:Hybrid retrieval can balance exact-code precision with semantic symptom recall in a single retrieval plan. Here the relevant capability must let the team search exact terms alongside semantic similarity. An alternative targeting match exact rare identifiers solves another problem.
Incorrect Answers
A:The workload requires both exact error identifiers and conceptual symptom matches. Hybrid search combines lexical retrieval with vector candidates, rather than using keywords alone.
B:Vector similarity helps conceptual recall but may lose exact identifiers. Hybrid retrieval combines code-sensitive lexical matching with symptom-oriented semantic search.
D:Semantic ranking operates on an existing candidate set and can miss exact product/error codes if keyword recall is omitted. Use a hybrid design that retains identifier matching.
Question 84
An employee asks about a procedure absent from every authorized document. What should a grounded chatbot do?
- Generate a likely procedure from related documents and present it as the official internal rule despite no matching authorized source.
- State that the available sources do not establish the requested procedure
- Attach source version and retrieval timestamp to citations
- Call the authoritative live inventory system for real-time figures
Explanation
Correct Answer
B:Explicit uncertainty is preferable to fabricating an internal policy not present in accessible evidence. This resolves the specific need to handle missing evidence without making up facts, while a measure that helps compute correct numbers from retrieved financial data addresses a different requirement.
Incorrect Answers
A:A plausible general practice is not evidence of the organization's procedure. The chatbot must acknowledge that its available authorized documents do not establish the requested rule.
C:Version provenance establishes whether the answer relied on current or historical reference material. This helps when a team needs to identify the precise evidence version; it does not establish how to handle missing evidence without making up facts.
D:A periodically updated index is not a reliable source for a fact that changes minute by minute. The function serves to ground answers on live operational data, whereas the actual requirement is to handle missing evidence without making up facts.
Question 85
Search results contain documents but the most useful passage is buried halfway through a 200-page manual. What ingestion change helps?
- Chunk source documents into coherent, indexed passages with source offsets
- Index every 200-page manual as one indivisible chunk and increase top-k, assuming that will pinpoint the useful passage.
- Schedule incremental ingestion with update and deletion tracking
- Rerank a single 200-page index document while leaving its internal chapters unsplit and unaddressable.
Explanation
Correct Answer
A:Passage granularity improves retrieval specificity while source offsets preserve provenance for citations. This directly supports the need to retrieve useful subsections of long documents, unlike a mechanism primarily intended to index scanned text while preserving table context.
Incorrect Answers
B:More whole-manual hits cannot isolate the answer-bearing section. Coherent chunks with stable source offsets let retrieval return the relevant procedure without flooding context.
C:Index freshness depends on propagating changes and removals from authoritative sources into searchable content. Here the needed outcome is to retrieve useful subsections of long documents; an option designed to keep indexed content synchronized with updates would not directly provide it.
D:Ranking a whole manual doesn't create passage-level retrieval boundaries. Split the document into coherent chunks with source offsets before retrieval and optional reranking.
Question 86
An assistant generates citations that point to the right file but the wrong paragraph. Which output check is missing?
- Link every claim only to the document title without locating any paragraph.
- Verify each cited span maps to retrieved source text supporting the corresponding claim.
- Check that the cited document exists but not that the cited span matches.
- Allow the model to fabricate page numbers if a file was retrieved.
Explanation
Correct Answer
B:Citations need passage-level provenance, not merely plausible document titles or URLs.
Incorrect Answers
A:A document-level title is too broad to demonstrate support for an individual assertion.
C:Existence of a source cannot establish that a particular paragraph contains supporting evidence.
D:Invented page references break traceability and are not grounded citation verification.
Question 87
A retrieval system returns relevant documents but results are poorly ordered by actual user intent. What ranking capability may improve top hits?
- Evaluate semantic ranking on the initial candidate set
- Keep results in original ingestion order after candidate retrieval even when the user's question targets a specific procedure within a long manual.
- Force an exact product-code match to the top of every search, even when the user's query asks for a different intent.
- Combine lexical and vector queries in hybrid search
Explanation
Correct Answer
A:Semantic reranking uses richer intent matching after initial retrieval to improve which candidates are shown first. The decisive difference is whether the design can rerank relevant candidates for question fit; the competing function serves to retrieve by semantic meaning rather than exact words.
Incorrect Answers
B:Retrieval order does not guarantee that the most useful passage appears first. A semantic reranker can reassess candidates against the query's actual information need.
C:Identifier equality alone doesn't reorder generally relevant results according to meaning. Evaluate semantic ranking or appropriate hybrid result fusion against labeled queries.
D:Hybrid retrieval can balance exact-code precision with semantic symptom recall in a single retrieval plan. That is relevant to search exact terms alongside semantic similarity, but the present scenario requires an ability to rerank relevant candidates for question fit.
Question 88
Two current retrieved policies disagree about the same requirement. What response should the system produce?
- Choose the policy document with the highest cosine similarity and silently present its provision as controlling despite the contradiction.
- State that the available sources do not establish the requested procedure
- Highlight the documented conflict and route the issue to the authorized policy owner
- Apply tenant and document-access filters before passing retrieved passages to the model
Explanation
Correct Answer
C:Conflicting authoritative passages require resolution, not silent selection of whichever rank is highest. The practical requirement is to reconcile contradictory retrieved policies transparently; a solution built to compute correct numbers from retrieved financial data would not meet that same need.
Incorrect Answers
A:Similarity ranking is not a policy-authority rule. The response must identify the documented conflict and escalate to the designated policy owner rather than invent a resolution.
B:Explicit uncertainty is preferable to fabricating an internal policy not present in accessible evidence. This helps when a team needs to handle missing evidence without making up facts; it does not establish how to reconcile contradictory retrieved policies transparently.
D:Retrieval relevance must remain inside the caller's authorized tenant scope. That capability is appropriate for prevent cross-tenant evidence mixing, not this separate task of finding how to reconcile contradictory retrieved policies transparently.
Question 89
A batch pipeline moves between two Foundry projects and suddenly receives 404 for its deployment name. What configuration should be compared?
- Reuse the previous project's endpoint while changing only a local display name.
- Rotate storage account keys to repair the model deployment not-found response.
- Validate that the client uses the target project's endpoint and the deployed model identifier in that project.
- Assume every project shares the same deployment namespace and name.
Explanation
Correct Answer
C:A deployment identifier is resolved in the appropriate project or resource context; the same friendly name is not necessarily deployed in every environment.
Incorrect Answers
A:The call may target the wrong service context, where the named deployment does not exist.
B:Storage keys do not fix a mismatched Foundry project endpoint or model deployment identifier.
D:Deployments may differ by project or resource; identical names are not automatically available in both.
Question 90
A search result has a matching heading but its body discusses a different product. What should the grounding layer do?
- Accept the heading match as sufficient evidence and cite the snippet without examining the body text or product identifier.
- Filter or rerank using substantive passage content rather than title matches alone
- Highlight the documented conflict and route the issue to the authorized policy owner
- Attach source version and retrieval timestamp to citations
Explanation
Correct Answer
B:The useful evidence must actually answer the question, not merely contain similar headline words. Here the relevant capability must let the team reject superficial result-title matches. An alternative targeting compute correct numbers from retrieved financial data solves another problem.
Incorrect Answers
A:Headings can be duplicated across product families. Grounding should inspect the substantive matched passage and its product context before accepting it as evidence for the user's request.
C:Conflicting authoritative passages require resolution, not silent selection of whichever rank is highest. That capability is appropriate for reconcile contradictory retrieved policies transparently, not this separate task of finding how to reject superficial result-title matches.
D:Version provenance establishes whether the answer relied on current or historical reference material. Here the needed outcome is to reject superficial result-title matches; an option designed to identify the precise evidence version would not directly provide it.