Question 140
A company must block a short list of prohibited product names, even if a general harm classifier rates the sentence harmless. Which control is appropriate?
- Review the deployment's supported content-filter thresholds and policy assignments
- Configure the exact organizational terms in a supported custom blocklist
- Queue the proposed decision for an authorized human reviewer before committing it
- Apply Prompt Shields to detect suspicious instructions embedded in retrieved content
Explanation
Correct Answer
B:A term-specific blocklist enforces business policy independently of broad harm categories. In the public generative-AI assistant, confirm the fix by retesting the affected behavior after configuring the exact organizational terms in a supported custom blocklist.
Incorrect Answers
A:Moderation behavior depends on configured content controls; changing token limits would not enforce harm-category policy in its own workflow, but that is a different correction from the present requirement to configure the exact organizational terms in a supported custom blocklist. The reported issue would remain.
C:That change addresses another failure mode: human approval makes high-impact decisions reviewable; presenting a confidence score alone is not accountability; it does not address this case, for which the relevant intervention is to configure the exact organizational terms in a supported custom blocklist.
D:Although useful elsewhere in the public generative-AI assistant, indirect prompt injection attempts to turn lower-trust source text into instructions, distinct from ordinary malicious wording in a direct user prompt; the present case is different and calls for the team to configure the exact organizational terms in a supported custom blocklist.
Question 141
A release for a private-document answer assistant is blocked until the application can let a user trace an AI answer back to the retrieved documents used as support. The tester will verify that the answer identifies the supporting source records. What should be implemented?
- Attach source metadata and citations to answers
- Generate embeddings for document chunks
- Augment the model prompt with retrieved evidence
- Chunk source documents into passages
Explanation
Correct Answer
A:Citation-aware output carries source identifiers, titles, or links forward from retrieval into the response. For this private-document answer assistant, the intended evidence is that the answer identifies the supporting source records. Citations are not created merely by computing embeddings if source metadata is discarded.
Incorrect Answers
B:An embedding model maps text into a vector space where related meaning can be compared mathematically; it may be used to encode each passage as a numeric vector for semantic similarity in a private-document answer assistant, but the current requirement is instead to let a user trace an AI answer back to the retrieved documents used as support. That separate capability does not prove that the answer identifies the supporting source records; Embeddings are not natural-language answers and do not enforce access to private documents, so selecting that alternative would not meet the stated acceptance check.
C:The augmentation step puts retrieved context into the model input so generation can use current enterprise data; it may be used to provide the selected source passages alongside the question and instructions in a private-document answer assistant, but the current requirement is instead to let a user trace an AI answer back to the retrieved documents used as support. That separate capability does not prove that the answer identifies the supporting source records; It does not update the model's pretrained weights or replace the need for retrieval, so selecting that alternative would not meet the stated acceptance check.
D:Chunking produces passages sized for retrieval and model context; metadata can preserve document lineage; it may be used to split lengthy source documents into retrievable sections with appropriate overlap in a private-document answer assistant, but the current requirement is instead to let a user trace an AI answer back to the retrieved documents used as support. That separate capability does not prove that the answer identifies the supporting source records; It does not by itself convert passages into semantic vectors or compose final answers, so selecting that alternative would not meet the stated acceptance check.
Question 142
Documents exist in the search index but the retriever returns no passage for a very specific phrase. What should be inspected?
- Check indexed searchable fields, query formulation, and retrieval configuration
- Require a supported-answer fallback when the grounding set is empty or insufficient
- Reduce chunk size before embedding and preserve document-to-chunk identifiers
- Record source version or retrieval timestamp alongside the cited document identifier
Explanation
Correct Answer
A:Content cannot be retrieved when it was not indexed into searchable fields or the request addresses the wrong index/field. In the private-document answer assistant, confirm the fix by retesting the affected behavior after checking indexed searchable fields, query formulation, and retrieval configuration.
Incorrect Answers
B:That change addresses another failure mode: rAG should not manufacture facts in the absence of evidence; an explicit abstention path is safer; it does not address this case, for which the relevant intervention is to check indexed searchable fields, query formulation, and retrieval configuration.
C:Although useful elsewhere in the private-document answer assistant, embedding input limits are model constraints; truncating oversized passages silently may discard important evidence; the present case is different and calls for the team to check indexed searchable fields, query formulation, and retrieval configuration.
D:Versioned provenance helps determine which source text supported an earlier answer in its own workflow, but that is a different correction from the present requirement to check indexed searchable fields, query formulation, and retrieval configuration. The reported issue would remain.
Question 143
The answer assistant finds a relevant policy chunk, but only after thousands of unrelated hits. How can retrieval precision improve?
- Apply supported metadata filters such as tenant and document validity before grounding
- Keep stable document IDs and page or chunk locators in retrieval results
- Inject the selected evidence passages into the model request context
- Use compatible embedding models and dimensions for indexed documents and query inputs
Explanation
Correct Answer
A:Filtering the candidate corpus by structured constraints reduces irrelevant or unauthorized retrieval results. In the private-document answer assistant, confirm the fix by retesting the affected behavior after applying supported metadata filters such as tenant and document validity before grounding.
Incorrect Answers
B:Although useful elsewhere in the private-document answer assistant, fine-grained source locators make generated citations navigable and auditable; the present case is different and calls for the team to apply supported metadata filters such as tenant and document validity before grounding.
C:That change addresses another failure mode: retrieval does not ground generation until the passages are actually supplied to the model's input; it does not address this case, for which the relevant intervention is to apply supported metadata filters such as tenant and document validity before grounding.
D:Vector comparison requires schema-compatible vectors in the same intended representation space in its own workflow, but that is a different correction from the present requirement to apply supported metadata filters such as tenant and document validity before grounding. The reported issue would remain.
Question 144
The model answers confidently when retrieval returns no relevant passages. What safety behavior should the prompt and app implement?
- Check indexed searchable fields, query formulation, and retrieval configuration
- Reduce chunk size before embedding and preserve document-to-chunk identifiers
- Record source version or retrieval timestamp alongside the cited document identifier
- Require a supported-answer fallback when the grounding set is empty or insufficient
Explanation
Correct Answer
D:RAG should not manufacture facts in the absence of evidence; an explicit abstention path is safer. In the private-document answer assistant, confirm the fix by retesting the affected behavior after requiring a supported-answer fallback when the grounding set is empty or insufficient.
Incorrect Answers
A:Although useful elsewhere in the private-document answer assistant, content cannot be retrieved when it was not indexed into searchable fields or the request addresses the wrong index/field; the present case is different and calls for the team to require a supported-answer fallback when the grounding set is empty or insufficient.
B:That change addresses another failure mode: embedding input limits are model constraints; truncating oversized passages silently may discard important evidence; it does not address this case, for which the relevant intervention is to require a supported-answer fallback when the grounding set is empty or insufficient.
C:Versioned provenance helps determine which source text supported an earlier answer in its own workflow, but that is a different correction from the present requirement to require a supported-answer fallback when the grounding set is empty or insufficient. The reported issue would remain.
Question 145
A search pipeline generates vectors for documents, but query vectors use an unrelated embedding model with different dimensions. What is the fix?
- Apply supported metadata filters such as tenant and document validity before grounding
- Use compatible embedding models and dimensions for indexed documents and query inputs
- Keep stable document IDs and page or chunk locators in retrieval results
- Split the policy into separately retrievable passages with source-page metadata
Explanation
Correct Answer
B:Vector comparison requires schema-compatible vectors in the same intended representation space. In the private-document answer assistant, confirm the fix by retesting the affected behavior after using compatible embedding models and dimensions for indexed documents and query inputs.
Incorrect Answers
A:Filtering the candidate corpus by structured constraints reduces irrelevant or unauthorized retrieval results in its own workflow, but that is a different correction from the present requirement to use compatible embedding models and dimensions for indexed documents and query inputs. The reported issue would remain.
C:Although useful elsewhere in the private-document answer assistant, fine-grained source locators make generated citations navigable and auditable; the present case is different and calls for the team to use compatible embedding models and dimensions for indexed documents and query inputs.
D:That change addresses another failure mode: chunked retrieval makes later portions independently searchable rather than representing an entire long file as one item; it does not address this case, for which the relevant intervention is to use compatible embedding models and dimensions for indexed documents and query inputs.
Question 146
A cited document was updated after the answer was generated. What provenance should the stored answer include?
- Interpret similarity as retrieval relevance rather than a guarantee that a passage is correct
- Reduce chunk size before embedding and preserve document-to-chunk identifiers
- Record source version or retrieval timestamp alongside the cited document identifier
- Check indexed searchable fields, query formulation, and retrieval configuration
Explanation
Correct Answer
C:Versioned provenance helps determine which source text supported an earlier answer. In the private-document answer assistant, confirm the fix by retesting the affected behavior after recording source version or retrieval timestamp alongside the cited document identifier.
Incorrect Answers
A:Embedding proximity estimates semantic relatedness, not the authenticity or correctness of a source claim in its own workflow, but that is a different correction from the present requirement to record source version or retrieval timestamp alongside the cited document identifier. The reported issue would remain.
B:Although useful elsewhere in the private-document answer assistant, embedding input limits are model constraints; truncating oversized passages silently may discard important evidence; the present case is different and calls for the team to record source version or retrieval timestamp alongside the cited document identifier.
D:That change addresses another failure mode: content cannot be retrieved when it was not indexed into searchable fields or the request addresses the wrong index/field; it does not address this case, for which the relevant intervention is to record source version or retrieval timestamp alongside the cited document identifier.
Question 147
The assistant lists invented source titles not present in search results. How should citation generation be constrained?
- Filter to the currently effective, access-authorized policy version
- Run embedding generation for the new content before expecting vector-neighbor retrieval
- Construct citations from actual retrieved source metadata rather than free-form model guesses
- Use structure-aware boundaries and suitable overlap between adjacent chunks
Explanation
Correct Answer
C:Source metadata is application evidence; unconstrained text generation can fabricate bibliographic details. In the private-document answer assistant, confirm the fix by retesting the affected behavior after constructing citations from actual retrieved source metadata rather than free-form model guesses.
Incorrect Answers
A:Although useful elsewhere in the private-document answer assistant, a semantic match can still be stale or out of scope; retrieval must honor business applicability; the present case is different and calls for the team to construct citations from actual retrieved source metadata rather than free-form model guesses.
B:That change addresses another failure mode: text fields alone do not create embeddings unless supported vectorization is configured; it does not address this case, for which the relevant intervention is to construct citations from actual retrieved source metadata rather than free-form model guesses.
D:Meaningful boundaries preserve local context; blindly fixed-length splitting can separate definitions from their qualifiers in its own workflow, but that is a different correction from the present requirement to construct citations from actual retrieved source metadata rather than free-form model guesses. The reported issue would remain.
Question 148
A team responsible for a private-document answer assistant needs to provide the selected source passages alongside the question and instructions; the acceptance check is that the model receives the same factual passages found by the retriever. Which capability should be selected?
- Chunk source documents into passages
- Generate embeddings for document chunks
- Augment the model prompt with retrieved evidence
- Attach source metadata and citations to answers
Explanation
Correct Answer
C:The augmentation step puts retrieved context into the model input so generation can use current enterprise data. For this private-document answer assistant, the intended evidence is that the model receives the same factual passages found by the retriever. It does not update the model's pretrained weights or replace the need for retrieval.
Incorrect Answers
A:Chunking produces passages sized for retrieval and model context; metadata can preserve document lineage; it may be used to split lengthy source documents into retrievable sections with appropriate overlap in a private-document answer assistant, but the current requirement is instead to provide the selected source passages alongside the question and instructions. That separate capability does not prove that the model receives the same factual passages found by the retriever; It does not by itself convert passages into semantic vectors or compose final answers, so selecting that alternative would not meet the stated acceptance check.
B:An embedding model maps text into a vector space where related meaning can be compared mathematically; it may be used to encode each passage as a numeric vector for semantic similarity in a private-document answer assistant, but the current requirement is instead to provide the selected source passages alongside the question and instructions. That separate capability does not prove that the model receives the same factual passages found by the retriever; Embeddings are not natural-language answers and do not enforce access to private documents, so selecting that alternative would not meet the stated acceptance check.
D:Citation-aware output carries source identifiers, titles, or links forward from retrieval into the response; it may be used to let a user trace an AI answer back to the retrieved documents used as support in a private-document answer assistant, but the current requirement is instead to provide the selected source passages alongside the question and instructions. That separate capability does not prove that the model receives the same factual passages found by the retriever; Citations are not created merely by computing embeddings if source metadata is discarded, so selecting that alternative would not meet the stated acceptance check.
Question 149
In a private-document answer assistant, engineers must split lengthy source documents into retrievable sections with appropriate overlap. The expected evidence is that search hits point to manageable source excerpts. Which implementation is the best match?
- Retrieve relevant passages from a search index
- Chunk source documents into passages
- Augment the model prompt with retrieved evidence
- Generate embeddings for document chunks
Explanation
Correct Answer
B:Chunking produces passages sized for retrieval and model context; metadata can preserve document lineage. For this private-document answer assistant, the intended evidence is that search hits point to manageable source excerpts. It does not by itself convert passages into semantic vectors or compose final answers.
Incorrect Answers
A:Retrieval queries a prepared index to select likely relevant documents using keyword, vector, or hybrid techniques; it may be used to find supporting content for a user's current question before generation in a private-document answer assistant, but the current requirement is instead to split lengthy source documents into retrievable sections with appropriate overlap. That separate capability does not prove that search hits point to manageable source excerpts; It does not guarantee the chat model will cite evidence unless the retrieved material is passed along and used, so selecting that alternative would not meet the stated acceptance check.
C:The augmentation step puts retrieved context into the model input so generation can use current enterprise data; it may be used to provide the selected source passages alongside the question and instructions in a private-document answer assistant, but the current requirement is instead to split lengthy source documents into retrievable sections with appropriate overlap. That separate capability does not prove that search hits point to manageable source excerpts; It does not update the model's pretrained weights or replace the need for retrieval, so selecting that alternative would not meet the stated acceptance check.
D:An embedding model maps text into a vector space where related meaning can be compared mathematically; it may be used to encode each passage as a numeric vector for semantic similarity in a private-document answer assistant, but the current requirement is instead to split lengthy source documents into retrievable sections with appropriate overlap. That separate capability does not prove that search hits point to manageable source excerpts; Embeddings are not natural-language answers and do not enforce access to private documents, so selecting that alternative would not meet the stated acceptance check.
Question 150
A RAG model receives two conflicting policy documents and chooses a rescinded version in its response. Which TWO corrections make the evidence more reliable? (Choose TWO.)
- Use structure-aware boundaries and suitable overlap between adjacent chunks
- Run embedding generation for the new content before expecting vector-neighbor retrieval
- State explicit source-priority and attribution rules when building the augmented prompt
- Construct citations from actual retrieved source metadata rather than free-form model guesses
- Filter the retrieved passages to authoritative currently effective documents
Explanation
Correct Answers
C:Source precedence and attribution help the model distinguish authoritative policy evidence from superseded or informal notes.
E:Grounding on valid source versions prevents the model from using stale material merely because it is semantically close to the question.
Incorrect Answers
A:Meaningful boundaries preserve local context; blindly fixed-length splitting can separate definitions from their qualifiers in its own workflow, but that is a different correction from the present requirement to filter the retrieved passages to authoritative currently effective documents and state explicit source-priority and attribution rules when building the augmented prompt. The reported issue would remain.
B:That change addresses another failure mode: text fields alone do not create embeddings unless supported vectorization is configured; it does not address this case, for which the relevant intervention is to filter the retrieved passages to authoritative currently effective documents and state explicit source-priority and attribution rules when building the augmented prompt.
D:Although useful elsewhere in the private-document answer assistant, source metadata is application evidence; unconstrained text generation can fabricate bibliographic details; the present case is different and calls for the team to filter the retrieved passages to authoritative currently effective documents and state explicit source-priority and attribution rules when building the augmented prompt.