Question 190
A PDF loader silently skips unreadable pages, yet marks the document fully indexed. What should the pipeline expose?
- Page-level failures and document completeness status
- A larger generation model for every later query
- A success flag whenever at least one page loads
- Only the number of successfully embedded chunks
Explanation
Correct Answer
A:Silent omission makes missing evidence look like a complete corpus. Explicit ingestion status allows remediation and prevents unsupported completeness claims.
Incorrect Answers
B:A model cannot retrieve pages that were never indexed. The upstream omission needs detection and repair in the ingestion process.
C:Partial success is not full ingestion. The status should distinguish a usable subset from a complete representation of the document.
D:Success counts do not reveal which expected pages were omitted. The pipeline needs a denominator and failure details for meaningful completeness.
Question 191
A user asks for the exact sum of approved invoices in a database. Text embeddings retrieve only a sample of invoice descriptions. Which integration is more appropriate?
- Sum amounts from the ten most similar descriptions
- An authorized structured query over the relevant records
- Ask the model to extrapolate the total from the sample
- Compute a database sum without filtering for approval status
Explanation
Correct Answer
B:Exact aggregation requires complete selection and deterministic arithmetic. Semantic similarity retrieval is not a reliable substitute for the database query conditions.
Incorrect Answers
A:A similarity sample does not establish that all qualifying invoices are included. The resulting total can omit valid records or include irrelevant ones.
C:Extrapolation produces an estimate, while the request requires an exact sum. The available structured source can support the exact calculation.
D:A database aggregate can still answer the wrong question if its selection is wrong. The exact requested total must include only the authorized approved invoices.
Question 192
An analyst asks which suppliers are indirectly connected through shared owners across three levels. Which data access pattern best fits explicit ownership relationships?
- Retrieve documents whose wording resembles shared owners
- Traverse the authorized relationship graph with bounded depth
- Look up only each supplier's immediate owner without following further links
- Rank suppliers by document popularity without traversing ownership links
Explanation
Correct Answer
B:The question concerns paths between entities, not just textual similarity. Graph traversal can apply the specified relationship type and depth systematically.
Incorrect Answers
A:Semantic search may locate useful descriptions but does not guarantee complete multi-hop relationship traversal. It is a poor sole mechanism for this precise query.
C:A direct owner lookup covers one relationship level. The question requires indirect connections across three levels, so bounded traversal must follow the relevant intermediate relationships.
D:Popularity measures attention to records rather than shared ownership. It cannot answer the stated relationship query without using the relevant edges.
Question 193
A question compares retention rules for three document classes, but one broad search retrieves only the most common class. What should be tested?
- Retrieve one global top-k list without checking coverage of each required class
- Retrieve each required class separately and combine supported results
- Answer only for the class with the most passages
- Repeat the same broad search without changing its coverage
Explanation
Correct Answer
B:Decomposition makes the evidence requirements explicit for every comparison target. Synthesis can then preserve distinctions instead of relying on one dominant retrieval result.
Incorrect Answers
A:A global ranking can be dominated by one class and omit another. Separate coverage requirements should guide retrieval and synthesis for the multi-part request.
C:That silently narrows the user's comparison. The system should obtain the missing evidence or clearly state the limitation.
D:Repeated identical retrieval may return the same dominant class. The missing requirement is balanced evidence across the requested categories.
Question 194
A retriever combines lexical and vector scores by adding their raw values, although their ranges differ greatly. Why can this fail?
- One score scale can dominate without reflecting relevance
- Raw scores are directly comparable whenever both retrievers rank relevance
- Lexical and semantic retrieval can never be combined
- Increasing every score by the same constant fixes calibration
Explanation
Correct Answer
A:Raw magnitudes from different scoring systems are not automatically comparable. Fusion or calibration should be evaluated against relevance judgments.
Incorrect Answers
B:Different scoring functions can use unrelated scales even when both measure relevance. Adding raw values can let one numerical range dominate the combined ranking.
C:The methods can be complementary. The problem is an unjustified combination rule, not an inherent incompatibility of the retrieval signals.
D:A constant shift does not reconcile differing scales or distributions. The fusion design needs a meaningful relationship between the signals.
Question 195
Overlapping chunks from one document occupy all top retrieval positions and crowd out a second necessary source. What selection strategy should be evaluated?
- Apply the same relevance rank to every duplicated copy
- Increase the score of every duplicate chunk
- Discard the entire first document permanently
- Deduplicate or diversify candidates while preserving relevant coverage
Explanation
Correct Answer
D:Redundant near-identical passages can waste the context budget. Selection should balance strong evidence with coverage of distinct sources needed for the answer.
Incorrect Answers
A:Retaining identical high-ranked copies still crowds out distinct evidence. Candidate selection should account for redundancy as well as individual relevance.
B:This reinforces the crowding problem. Multiple copies of the same evidence do not replace a missing complementary source.
C:The document may contain necessary evidence. The goal is to reduce redundant representation, not remove a relevant source wholesale.
Question 196
An answer cites document ID 72, but reviewers cannot locate the supporting sentence in a long changing file. What metadata would most help?
- A document revision and stable passage locator
- Only the retrieval score of the selected passage
- Only the document's current filename
- The total number of characters in the source
Explanation
Correct Answer
A:A document identifier alone may refer to changing content. Version and location information let reviewers find the evidence that was actually supplied.
Incorrect Answers
B:A score describes matching strength under a particular retrieval process. It does not locate the cited content in a stable version of its source.
C:A filename can remain constant across edits and gives no passage location. It is insufficient for reconstructing a particular source claim.
D:Length is not a locator and can be shared by different revisions. It cannot reliably identify the cited passage.
Question 197
A support corpus uses English terminology, while users often search in French. Retrieval misses equivalent concepts even though access filters are correct. What experiment directly tests a retrieval-side remedy?
- Translate only the final generated answer after the same empty retrieval
- Remove document access filters to increase the number of candidates
- Increase the number of English-only examples in the answer-format prompt
- Evaluate meaning-preserving translation or multilingual query retrieval on labeled cross-language cases
Explanation
Correct Answer
D:The failure concerns matching concepts across languages. A controlled query transformation or compatible multilingual retrieval approach can be tested against known relevant passages.
Incorrect Answers
A:Output translation does not bring missing evidence into context. The retrieval stage still needs to locate the relevant English material.
B:The filters are stated to be correct and do not cause the language mismatch. Removing them would introduce unauthorized results without addressing the identified mechanism.
C:Answer-format examples act after retrieval and do not necessarily change query matching. The experiment should target the stage that misses cross-language evidence.
Question 198
A desktop application connects to two MCP servers. Which component should coordinate the user's model interaction and enforce the application's overall access policy?
- An individual MCP client managing only its own server connection
- Each server independently, without a policy owner in the desktop application
- The model alone, using the servers' tool descriptions as access rules
- The host application managing its MCP clients
Explanation
Correct Answer
D:The host owns the user-facing interaction and coordinates connections. Individual servers expose capabilities, but do not automatically govern the whole application's policy.
Incorrect Answers
A:Each client handles its connection, while the host coordinates the wider application experience and policy. No single connection alone has the stated application-wide role.
B:Servers enforce their own boundaries but do not replace the host's responsibility for the combined user interaction. The application still needs an overall policy coordinator.
C:Tool descriptions provide interface information, not an enforcement boundary. Application policy requires controls in the host and participating execution systems.
Question 199
A server needs to expose a reference document and a separate operation that updates a ticket. Which representation best reflects their roles?
- Two prompt templates that describe reading and updating the record
- A resource for reference content and a tool for the update
- Two document resources that both perform writes when opened
- A single unlabeled text blob containing both
Explanation
Correct Answer
B:The document is data to read, while the update is an executable operation. Distinguishing them clarifies how the client should access and authorize each capability.
Incorrect Answers
A:Templates can guide interaction but do not supply the underlying read or mutation capability. The integration needs the appropriate resource and executable tool contracts.
C:Reading reference content should not disguise a state-changing operation. Hidden side effects make authorization and user expectations difficult to enforce.
D:Unlabeled content hides the distinction between evidence and action. The client cannot reliably infer the update contract from arbitrary prose.
Question 200
A server supplies a reusable user-invoked workflow template through its prompt capability. What should a client avoid assuming?
- That a template can help structure repeated tasks
- That template inputs may need validation
- That the template grants permission to execute protected tools
- That the user can review a rendered template
Explanation
Correct Answer
C:A prompt can guide a workflow but does not confer access rights. Actual tool operations still require the application's and server's authorization checks.
Incorrect Answers
A:Reusable instructions can be useful for consistency. The limitation concerns authority, not the value of a well-designed template.
B:Inputs can influence rendered instructions and should be handled according to the contract. Validation remains relevant for reusable prompts.
D:Review can be appropriate before use. The template's existence does not prevent a client from making its content visible and understandable.