100% Latest & Updated Microsoft DP-800 Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 990
An application wants to check whether RAG helped over unguided prompts. The team selects: Use groundedness, relevance and citation-based evaluation. Which implementation action puts this design into practice?
- Use FOR JSON PATH or JSON_OBJECT projection of allowed fields
- Create holdout questions with known supporting passages
- Retrieve authorized current passages and pass them with question
- Apply RLS predicate and per-document ACL before prompt construction
Explanation
Correct Answer
B:Create holdout questions with known supporting passages. This is the concrete setup for use groundedness, relevance and citation-based evaluation, and compare hallucination rate and answer citation accuracy supplies a direct confirmation.
Incorrect Answers
A:Use FOR JSON PATH or JSON_OBJECT projection of allowed fields. This is relevant when a chatbot uses SQL result rows as context for a model. Instead, create holdout questions with known supporting passages.
C:Retrieve authorized current passages and pass them with question. This concerns an ai assistant answers questions from frequently changing internal procedures. Instead, create holdout questions with known supporting passages.
D:Apply RLS predicate and per-document ACL before prompt construction. This concerns rag that respects row-level security. Instead, create holdout questions with known supporting passages.
Question 991
An application wants to check whether RAG helped over unguided prompts. After implementing the design, which check most directly validates the expected behavior?
- Compare retrieved source IDs for different users
- Test contradictory version pair yields current rule
- Compare hallucination rate and answer citation accuracy
- Evaluate keyword and paraphrase cases in same test suite
Explanation
Correct Answer
C:Compare hallucination rate and answer citation accuracy. This verifies the requirement in the case rather than merely confirming that a job ran; an opposing result can reveal that no evaluated ground-truth and citation evidence exists.
Incorrect Answers
A:Compare retrieved source IDs for different users. This concerns rag that respects row-level security. Instead, compare hallucination rate and answer citation accuracy.
B:Test contradictory version pair yields current rule. This concerns a rag answer includes contradictory facts from old and new records. Instead, compare hallucination rate and answer citation accuracy.
D:Evaluate keyword and paraphrase cases in same test suite. This concerns hybrid SQL/vector candidates before generating answer. Instead, compare hallucination rate and answer citation accuracy.
Question 992
An application wants to check whether RAG helped over unguided prompts. A test instead finds: Team declares success from fluent demos only. Which underlying issue best explains this finding?
- No evaluated ground-truth and citation evidence exists
- Prompt lacks no-evidence behavior or model ignored it
- RAG pipeline bypasses SQL authorization
- Context included duplicate conflicting revisions
Explanation
Correct Answer
A:No evaluated ground-truth and citation evidence exists. It accounts for the specific observed failure: team declares success from fluent demos only. The intended solution is use groundedness, relevance and citation-based evaluation with the required setup.
Incorrect Answers
B:Prompt lacks no-evidence behavior or model ignored it. This concerns assistant confidently makes up non-existent product policy. Here, no evaluated ground-truth and citation evidence exists.
C:RAG pipeline bypasses SQL authorization. This concerns user sees facts from another tenant. Here, no evaluated ground-truth and citation evidence exists.
D:Context included duplicate conflicting revisions. This concerns model quotes obsolete value despite new approved record. Here, no evaluated ground-truth and citation evidence exists.
Question 993
A company wants hybrid SQL/vector candidates before generating answer. Which database design or engineering approach most directly meets this requirement?
- Fuse ranked full-text and semantic results then pass citations
- Validate and extract model JSON response before persistence
- Retrieve transactional current stock before text generation
- Define prompt template with instructions and bounded evidence section
Explanation
Correct Answer
A:Fuse ranked full-text and semantic results then pass citations. The implementation is generate hybrid top-k with rrf and bounded source context; success can be checked by evaluate keyword and paraphrase cases in same test suite. The method addresses the stated data and workload constraint.
Incorrect Answers
B:Validate and extract model JSON response before persistence. This is relevant when an assistant produces a json answer that must be stored in SQL. Here, fuse ranked full-text and semantic results then pass citations.
C:Retrieve transactional current stock before text generation. This concerns live stock levels from SQL. Here, fuse ranked full-text and semantic results then pass citations.
D:Define prompt template with instructions and bounded evidence section. This concerns a reusable prompt with SQL database retrieval. Here, fuse ranked full-text and semantic results then pass citations.
Question 994
A company wants hybrid SQL/vector candidates before generating answer. The team selects: Fuse ranked full-text and semantic results then pass citations. Which implementation action puts this design into practice?
- Generate hybrid top-K with RRF and bounded source context
- Retrieve authorized current passages and pass them with question
- Explicitly instruct refusal or escalation when context lacks answer
- Apply RLS predicate and per-document ACL before prompt construction
Explanation
Correct Answer
A:Generate hybrid top-K with RRF and bounded source context. This is the concrete setup for fuse ranked full-text and semantic results then pass citations, and evaluate keyword and paraphrase cases in same test suite supplies a direct confirmation.
Incorrect Answers
B:Retrieve authorized current passages and pass them with question. This concerns an ai assistant answers questions from frequently changing internal procedures. Instead, generate hybrid top-K with RRF and bounded source context.
C:Explicitly instruct refusal or escalation when context lacks answer. This concerns not invent answer when SQL retrieval returns no supporting rows. Instead, generate hybrid top-K with RRF and bounded source context.
D:Apply RLS predicate and per-document ACL before prompt construction. This concerns rag that respects row-level security. Instead, generate hybrid top-K with RRF and bounded source context.
Question 995
A company wants hybrid SQL/vector candidates before generating answer. After implementing the design, which check most directly validates the expected behavior?
- Test prompt injection phrase embedded in source document
- Test out-of-scope question returns grounded uncertainty
- Evaluate keyword and paraphrase cases in same test suite
- Test authorized response and telemetry
Explanation
Correct Answer
C:Evaluate keyword and paraphrase cases in same test suite. This verifies the requirement in the case rather than merely confirming that a job ran; an opposing result can reveal that retrieval uses only one of lexical or semantic signals.
Incorrect Answers
A:Test prompt injection phrase embedded in source document. This concerns a reusable prompt with SQL database retrieval. Instead, evaluate keyword and paraphrase cases in same test suite.
B:Test out-of-scope question returns grounded uncertainty. This concerns not invent answer when SQL retrieval returns no supporting rows. Instead, evaluate keyword and paraphrase cases in same test suite.
D:Test authorized response and telemetry. This concerns a stored procedure sends an allowed request to an external text model. Instead, evaluate keyword and paraphrase cases in same test suite.
Question 996
A company wants hybrid SQL/vector candidates before generating answer. A test instead finds: One class of queries consistently misses important evidence. Which underlying issue best explains this finding?
- Retrieval uses only one of lexical or semantic signals
- SQL rows were flattened without structured context
- Retrieved text was treated as higher-priority instructions
- Logging pipeline persists unredacted secrets
Explanation
Correct Answer
A:Retrieval uses only one of lexical or semantic signals. It accounts for the specific observed failure: one class of queries consistently misses important evidence. The intended solution is fuse ranked full-text and semantic results then pass citations with the required setup.
Incorrect Answers
B:SQL rows were flattened without structured context. This concerns llm receives ambiguous comma-separated values. Here, retrieval uses only one of lexical or semantic signals.
C:Retrieved text was treated as higher-priority instructions. This concerns assistant follows malicious instruction from retrieved data. Here, retrieval uses only one of lexical or semantic signals.
D:Logging pipeline persists unredacted secrets. This concerns prompt history stores connection strings verbatim. Here, retrieval uses only one of lexical or semantic signals.
Question 997
A compliance RAG app cannot expose secrets in query logs. Which database design or engineering approach most directly meets this requirement?
- Use groundedness, relevance and citation-based evaluation
- Fuse ranked full-text and semantic results then pass citations
- Redact sensitive prompts and protect model invocation logging
- Use sp_invoke_external_rest_endpoint where supported and configured
Explanation
Correct Answer
C:Redact sensitive prompts and protect model invocation logging. The implementation is configure audited redaction and role-based log access; success can be checked by inspect logs and telemetry for plain credentials. The method addresses the stated data and workload constraint.
Incorrect Answers
A:Use groundedness, relevance and citation-based evaluation. This concerns checking whether rag helped over unguided prompts. Here, redact sensitive prompts and protect model invocation logging.
B:Fuse ranked full-text and semantic results then pass citations. This concerns hybrid SQL/vector candidates before generating answer. Here, redact sensitive prompts and protect model invocation logging.
D:Use sp_invoke_external_rest_endpoint where supported and configured. This concerns a stored procedure sends an allowed request to an external text model. Here, redact sensitive prompts and protect model invocation logging.
Question 998
A compliance RAG app cannot expose secrets in query logs. The team selects: Redact sensitive prompts and protect model invocation logging. Which implementation action puts this design into practice?
- Configure audited redaction and role-based log access
- Explicitly instruct refusal or escalation when context lacks answer
- Configure endpoint address, credentials and model deployment
- Create SQL projection excluding secrets before external request
Explanation
Correct Answer
A:Configure audited redaction and role-based log access. This is the concrete setup for redact sensitive prompts and protect model invocation logging, and inspect logs and telemetry for plain credentials supplies a direct confirmation.
Incorrect Answers
B:Explicitly instruct refusal or escalation when context lacks answer. This concerns not invent answer when SQL retrieval returns no supporting rows. Instead, configure audited redaction and role-based log access.
C:Configure endpoint address, credentials and model deployment. This concerns connecting SQL retrieval to microsoft foundry model. Instead, configure audited redaction and role-based log access.
D:Create SQL projection excluding secrets before external request. This concerns a support bot routes answers to an external model via rest. Instead, configure audited redaction and role-based log access.
Question 999
A compliance RAG app cannot expose secrets in query logs. After implementing the design, which check most directly validates the expected behavior?
- Compare hallucination rate and answer citation accuracy
- Test authorized response and telemetry
- Inspect logs and telemetry for plain credentials
- Inspect outbound JSON for PII and hidden columns
Explanation
Correct Answer
C:Inspect logs and telemetry for plain credentials. This verifies the requirement in the case rather than merely confirming that a job ran; an opposing result can reveal that logging pipeline persists unredacted secrets.
Incorrect Answers
A:Compare hallucination rate and answer citation accuracy. This concerns checking whether rag helped over unguided prompts. Instead, inspect logs and telemetry for plain credentials.
B:Test authorized response and telemetry. This concerns a stored procedure sends an allowed request to an external text model. Instead, inspect logs and telemetry for plain credentials.
D:Inspect outbound JSON for PII and hidden columns. This concerns a support bot routes answers to an external model via rest. Instead, inspect logs and telemetry for plain credentials.
Question 1000
A compliance RAG app cannot expose secrets in query logs. A test instead finds: Prompt history stores connection strings verbatim. Which underlying issue best explains this finding?
- RAG pipeline bypasses SQL authorization
- RAG context construction over-shared database columns
- Logging pipeline persists unredacted secrets
- Retrieval uses only one of lexical or semantic signals
Explanation
Correct Answer
C:Logging pipeline persists unredacted secrets. It accounts for the specific observed failure: prompt history stores connection strings verbatim. The intended solution is redact sensitive prompts and protect model invocation logging with the required setup.
Incorrect Answers
A:RAG pipeline bypasses SQL authorization. This concerns user sees facts from another tenant. Here, logging pipeline persists unredacted secrets.
B:RAG context construction over-shared database columns. This concerns external model receives unrelated customer details. Here, logging pipeline persists unredacted secrets.
D:Retrieval uses only one of lexical or semantic signals. This concerns one class of queries consistently misses important evidence. Here, logging pipeline persists unredacted secrets.
Developing AI-Enabled Database Solutions (DP-800) Practice Tests