cert
cert-1
cert-2

Pass Microsoft AI-300 Exam in First Attempt Guaranteed!

Get 100% Latest Exam Questions, Accurate & Verified Answers to Pass the Actual Exam!
30 Days Free Updates, Instant Download!

cert-5
cert-6
AI-300 Exam - Verified By Experts
AI-300 Premium File

AI-300 Premium File

$69.99
$76.99
  • Premium File 213 Questions & Answers. Last Update: Nov 12, 2026

Whats Included:

  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates
 
$76.99
$69.99
accept 45 downloads in the last 7 days
block-screenshots
Microsoft AI-300 Practice Test 60

Question 590

While optimizing a production RAG or fine-tuned model service, the following condition occurs: An enterprise fine-tuned assistant passes its original launch test but performs poorly after new product codes appear, and no evaluation monitor samples the updated terminology. The resulting model must meet held-out quality and security criteria. Which TWO complementary choices identify the issue and correct it? Select TWO.

  1. Review live prompt distribution, new product identifier coverage, task quality metrics, fine-tuned model version and evaluation refresh cadence.
  2. Inspect registered fine-tuned model version, retained checkpoints, base model lineage, tokenizer files, serving configuration and retention policy.
  3. Calibrate retrieval threshold and ranking settings using labeled queries to admit relevant context without overwhelming the model with noise.
  4. Refresh a governed evaluation set with recent terminology and trigger controlled retraining only after measuring sustained quality regression.
  5. Rebalance curated synthetic and real cases to preserve rare operational scenarios and evaluate on an untouched real-world edge-case set.

Explanation

Correct Answers

A:

Review live prompt distribution, new product identifier coverage, task quality metrics, fine-tuned model version and evaluation refresh cadence. This is the most relevant evidence for fine-tune evaluation lacks drift checks for new customer terminology. Fine-tuned models can become stale as input distribution shifts; continuous quality monitoring is needed beyond deployment-day accuracy. The matching controlled remediation is: Refresh a governed evaluation set with recent terminology and trigger controlled retraining only after measuring sustained quality regression. This matters here because while optimizing a production RAG or fine-tuned model service, the following condition occurs: An enterprise….

D:

Refresh a governed evaluation set with recent terminology and trigger controlled retraining only after measuring sustained quality regression. This targets fine-tune evaluation lacks drift checks for new customer terminology. Fine-tuned models can become stale as input distribution shifts; continuous quality monitoring is needed beyond deployment-day accuracy. Verify the failure through: Review live prompt distribution, new product identifier coverage, task quality metrics, fine-tuned model version and evaluation refresh cadence. Validate success with: Previously unseen product queries are monitored and subsequent vetted versions meet current quality targets without erasing old task competence. This matters here because while optimizing a production RAG or fine-tuned model service, the following condition occurs: An enterprise….

Incorrect Answers

B:

Inspect registered fine-tuned model version, retained checkpoints, base model lineage, tokenizer files, serving configuration and retention policy. This is associated with fine-tune production endpoint references deleted training artifact, not the essential evidence-and-fix pair for fine-tune evaluation lacks drift checks for new customer terminology. First examine: Review live prompt distribution, new product identifier coverage, task quality metrics, fine-tuned model version and evaluation refresh cadence. Then correct: Refresh a governed evaluation set with recent terminology and trigger controlled retraining only after measuring sustained quality regression.

C:

Calibrate retrieval threshold and ranking settings using labeled queries to admit relevant context without overwhelming the model with noise. This is associated with rag similarity threshold filters all relevant evidence, not the essential evidence-and-fix pair for fine-tune evaluation lacks drift checks for new customer terminology. First examine: Review live prompt distribution, new product identifier coverage, task quality metrics, fine-tuned model version and evaluation refresh cadence. Then correct: Refresh a governed evaluation set with recent terminology and trigger controlled retraining only after measuring sustained quality regression.

E:

Rebalance curated synthetic and real cases to preserve rare operational scenarios and evaluate on an untouched real-world edge-case set. This is associated with synthetic data overwhelms minority real-world edge cases, not the essential evidence-and-fix pair for fine-tune evaluation lacks drift checks for new customer terminology. First examine: Review live prompt distribution, new product identifier coverage, task quality metrics, fine-tuned model version and evaluation refresh cadence. Then correct: Refresh a governed evaluation set with recent terminology and trigger controlled retraining only after measuring sustained quality regression.

Question 591

During a Foundry generative AI quality and observability incident, the following condition occurs: A tool-enabled finance agent passes harmful-text screening but executes instructions inside a hostile retrieved web page because its safety suite contains no prompt-injection or tool-misuse cases. The fix must preserve evaluation validity and avoid unnecessary user-data collection. Which observable postchange result proves the intended remediation succeeded?

  1. An intentionally degraded candidate fails before production promotion and a passing candidate proceeds with an auditable evaluation receipt.
  2. The pipeline correctly flags answers that fail the user's original task even when they match an intermediate search query.
  3. The monitoring dataset reflects significant live workflows and reveals failures in the previously omitted segments.
  4. An adversarial document cannot cause unapproved transactions while authorized research questions still succeed.

Explanation

Correct Answer

D:

An adversarial document cannot cause unapproved transactions while authorized research questions still succeed. Harmful-content detection alone does not test whether an agent obeys malicious instructions inserted into lower-trust retrieved material. a direct corrective action is to add approved prompt-injection and tool-abuse evaluation cases with tool permission gates and block release on unsafe execution. This matters here because a tool-enabled finance agent passes harmful-text screening but executes instructions inside a hostile retrieved web….

Incorrect Answers

A:

An intentionally degraded candidate fails before production promotion and a passing candidate proceeds with an auditable evaluation receipt. That addresses a different configuration issue rather than the main requirement. Verify instead: An adversarial document cannot cause unapproved transactions while authorized research questions still succeed.

B:

The pipeline correctly flags answers that fail the user's original task even when they match an intermediate search query. That addresses a different configuration issue rather than the main requirement. Verify instead: An adversarial document cannot cause unapproved transactions while authorized research questions still succeed.

C:

The monitoring dataset reflects significant live workflows and reveals failures in the previously omitted segments. That addresses a different configuration issue rather than the main requirement. Verify instead: An adversarial document cannot cause unapproved transactions while authorized research questions still succeed.

Question 592

During a Microsoft Foundry generative AI deployment review, the following condition occurs: A team claims prompt B is better than prompt A, but B runs on a newer foundation model so the measured improvement cannot be attributed to the prompt change. The deployment must retain governed identities, endpoint contracts and auditable release history. Which observable postchange result proves the intended remediation succeeded?

  1. Representative order queries produce parseable permitted fields and reject or safely handle malformed outputs.
  2. Prompt tests retain representative behavior without exposing customer identities in accessible Git history or future deployments.
  3. Requests from the approved VNet resolve the private address and succeed, while public connectivity remains blocked.
  4. The comparison isolates prompt differences and produces reproducible evidence of quality and latency trade-offs.

Explanation

Correct Answer

D:

The comparison isolates prompt differences and produces reproducible evidence of quality and latency trade-offs. Experimental prompt comparisons require controlling confounders; a simultaneous model upgrade prevents a valid prompt-only attribution. a direct corrective action is to evaluate both prompt variants against the same pinned model and representative dataset with controlled assignment and consistent metrics. This matters here because a team claims prompt B is better than prompt A, but B runs on a newer foundation model so the measured improvement….

Incorrect Answers

A:

Representative order queries produce parseable permitted fields and reject or safely handle malformed outputs. That addresses a different configuration issue rather than the main requirement. Verify instead: The comparison isolates prompt differences and produces reproducible evidence of quality and latency trade-offs.

B:

Prompt tests retain representative behavior without exposing customer identities in accessible Git history or future deployments. That addresses a different configuration issue rather than the main requirement. Verify instead: The comparison isolates prompt differences and produces reproducible evidence of quality and latency trade-offs.

C:

Requests from the approved VNet resolve the private address and succeed, while public connectivity remains blocked. That addresses a different configuration issue rather than the main requirement. Verify instead: The comparison isolates prompt differences and produces reproducible evidence of quality and latency trade-offs.

Question 593

While optimizing a production RAG or fine-tuned model service, the following condition occurs: A synthetic data workflow for complaint classification injects real support tickets verbatim as prompts and emits generated examples still containing customer names and account details. The resulting model must meet held-out quality and security criteria. What should the engineer examine first to confirm the specific AI-300 workload issue?

  1. Compare synthetic-to-real sampling weights, rare-class coverage, holdout error breakdown and generated example diversity.
  2. Inspect information update frequency, retrieval freshness requirement, model task errors, training cost and data-change workflow.
  3. Examine synthetic-generation input provenance, PII detection scans, privacy transformations, output similarity to real records and review approvals.
  4. Examine source document structure, chunk boundaries, overlap size, retrieved context, query examples and groundedness errors.

Explanation

Correct Answer

C:

Examine synthetic-generation input provenance, PII detection scans, privacy transformations, output similarity to real records and review approvals. This is the most relevant evidence for synthetic training examples copy private customer identifiers. Generated data is not automatically privacy-safe; a model can reproduce sensitive source details when the generation prompt includes them. The matching controlled remediation is: Generate sanitized synthetic patterns from approved nonidentifying templates and screen outputs for memorized or copied private content.

Incorrect Answers

A:

Compare synthetic-to-real sampling weights, rare-class coverage, holdout error breakdown and generated example diversity. This would investigate synthetic data overwhelms minority real-world edge cases rather than the documented synthetic training examples copy private customer identifiers issue. The relevant investigation is: Examine synthetic-generation input provenance, PII detection scans, privacy transformations, output similarity to real records and review approvals.

B:

Inspect information update frequency, retrieval freshness requirement, model task errors, training cost and data-change workflow. This would investigate fine-tuning task selected when retrieval would better handle changing facts rather than the documented synthetic training examples copy private customer identifiers issue. The relevant investigation is: Examine synthetic-generation input provenance, PII detection scans, privacy transformations, output similarity to real records and review approvals.

D:

Examine source document structure, chunk boundaries, overlap size, retrieved context, query examples and groundedness errors. This would investigate rag retrieval chunks truncate critical policy exception rather than the documented synthetic training examples copy private customer identifiers issue. The relevant investigation is: Examine synthetic-generation input provenance, PII detection scans, privacy transformations, output similarity to real records and review approvals.

Question 594

During a Foundry generative AI quality and observability incident, the following condition occurs: An incident team enables verbose AI agent tracing that captures full API keys and customer documents into a broadly accessible Log Analytics workspace. The fix must preserve evaluation validity and avoid unnecessary user-data collection. Which acceptance test shows the AI operations defect is corrected?

  1. The responsible stage is identified and P95 user response time improves without gratuitous changes to healthy components.
  2. Authorized engineers can reproduce request failures without finding credentials or customer PII in log exports.
  3. Multilingual adversarial probes are handled consistently within the approved safety policy and ordinary requests remain usable.
  4. The report reconciles billed service usage with per-feature consumption and unnecessary repeat embedding calls decline.

Explanation

Correct Answer

B:

Authorized engineers can reproduce request failures without finding credentials or customer PII in log exports. Observability data has its own privacy and security boundary; unrestricted prompt and tool payload logging can create new exposure risks. a direct corrective action is to mask sensitive tokens and personal data, apply least-privilege log access and retain only necessary structured debugging evidence. This matters here because an incident team enables verbose AI agent tracing that captures full API keys and customer documents into a broadly….

Incorrect Answers

A:

The responsible stage is identified and P95 user response time improves without gratuitous changes to healthy components. That addresses a different configuration issue rather than the main requirement. Verify instead: Authorized engineers can reproduce request failures without finding credentials or customer PII in log exports.

C:

Multilingual adversarial probes are handled consistently within the approved safety policy and ordinary requests remain usable. That addresses a different configuration issue rather than the main requirement. Verify instead: Authorized engineers can reproduce request failures without finding credentials or customer PII in log exports.

D:

The report reconciles billed service usage with per-feature consumption and unnecessary repeat embedding calls decline. That addresses a different configuration issue rather than the main requirement. Verify instead: Authorized engineers can reproduce request failures without finding credentials or customer PII in log exports. This matters here because an incident team enables verbose AI agent tracing that captures full API keys and customer documents into a broadly….

Question 595

While deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: A central model team registers a reusable component in a shared Azure ML registry, but application workspace users cannot resolve the component because registry access was never granted. The change must remain limited to authorized workspaces and avoid unreviewed production impact. Which production-safe remediation should the engineer apply to this AI workflow?

  1. Replace embedded workspace-specific paths with parameterized, typed data inputs and bind them explicitly in each consuming pipeline.
  2. Give the approved compute or workspace identity the necessary ACR pull access and use a validated immutable image reference.
  3. Grant controlled registry consumption rights and use the fully qualified versioned registry asset reference in the target workspace.
  4. Require protected pull requests and automated component tests before changes to production training code are merged.

Explanation

Correct Answer

C:

Grant controlled registry consumption rights and use the fully qualified versioned registry asset reference in the target workspace. This targets registry model asset not available to second workspace. Cross-workspace registries support reusing vetted assets but do not automatically authorize every workspace or resolve unknown versions. Verify the failure through: Check the component's registry name and version, target workspace identity, registry read permissions and referenced environment dependencies. Validate success with: Authorized workspaces run the shared component while unauthorized subscriptions remain blocked from the registry.

Incorrect Answers

A:

Replace embedded workspace-specific paths with parameterized, typed data inputs and bind them explicitly in each consuming pipeline. This could address reusable component hardcodes workspace-specific asset path, but it does not solve the specified registry model asset not available to second workspace failure. The matching remediation is: Grant controlled registry consumption rights and use the fully qualified versioned registry asset reference in the target workspace.

B:

Give the approved compute or workspace identity the necessary ACR pull access and use a validated immutable image reference. This could address docker environment image cannot pull from private registry, but it does not solve the specified registry model asset not available to second workspace failure. The matching remediation is: Grant controlled registry consumption rights and use the fully qualified versioned registry asset reference in the target workspace.

D:

Require protected pull requests and automated component tests before changes to production training code are merged. This could address notebook merge bypasses protected evaluation tests, but it does not solve the specified registry model asset not available to second workspace failure. The matching remediation is: Grant controlled registry consumption rights and use the fully qualified versioned registry asset reference in the target workspace.

Question 596

While preparing a secure Foundry model or agent release, the following condition occurs: A GenAI team commits real customer tickets containing identifiers as few-shot prompt examples in a shared Git repository, exposing data beyond the approved processing boundary. The remediation must meet enterprise access and reliability requirements without a broad bypass. Which acceptance test shows the AI operations defect is corrected?

  1. The comparison isolates prompt differences and produces reproducible evidence of quality and latency trade-offs.
  2. Quality improves on relevant cohorts without exposing personal information or degrading approved common-case answers.
  3. Prompt tests retain representative behavior without exposing customer identities in accessible Git history or future deployments.
  4. Replaying approved test prompts yields compatible output contracts until the next reviewed version is promoted.

Explanation

Correct Answer

C:

Prompt tests retain representative behavior without exposing customer identities in accessible Git history or future deployments. Source control is not an appropriate unreviewed store for sensitive customer data; prompt versioning requires privacy-preserving examples and access governance. a direct corrective action is to remove direct identifiers from ongoing prompt examples, rotate affected credentials if necessary and use sanitized synthetic cases under review.

Incorrect Answers

A:

The comparison isolates prompt differences and produces reproducible evidence of quality and latency trade-offs. That addresses a different configuration issue rather than the main requirement. Verify instead: Prompt tests retain representative behavior without exposing customer identities in accessible Git history or future deployments.

B:

Quality improves on relevant cohorts without exposing personal information or degrading approved common-case answers. That addresses a different configuration issue rather than the main requirement. Verify instead: Prompt tests retain representative behavior without exposing customer identities in accessible Git history or future deployments.

D:

Replaying approved test prompts yields compatible output contracts until the next reviewed version is promoted. That addresses a different configuration issue rather than the main requirement. Verify instead: Prompt tests retain representative behavior without exposing customer identities in accessible Git history or future deployments.

Question 597

During a retrieval or model-customization performance review, the following condition occurs: A customer support model fine-tuned on historical chats becomes inconsistent because examples mix strict JSON tool calls with casual prose for identical task types. The change must keep citations, privacy and previously approved behavior intact. Which targeted Azure AI operations change best addresses the documented problem?

  1. Retain governed immutable fine-tune artifacts and record deployment lineage with tested restore/rollback procedures.
  2. Use an appropriate hybrid keyword-plus-vector search strategy with evaluated fusion and optional semantic reranking.
  3. Clean and stratify the supervised fine-tuning data to a coherent supported task format before retraining and evaluating.
  4. Halt promotion, review training data and fine-tuning choices, and require safety regression gates on the customized candidate before release.

Explanation

Correct Answer

C:

Clean and stratify the supervised fine-tuning data to a coherent supported task format before retraining and evaluating. This targets supervised fine-tuning dataset includes conflicting assistant behaviors. Supervised fine-tuning learns patterns in labeled examples; contradictory targets can degrade reliability even when many examples are provided. Verify the failure through: Inspect fine-tuning records, role/message formatting, target response schemas, label consistency and curated holdout behavior. Validate success with: The customized model returns the approved structured behavior on held-out scenarios without a spike in schema violations.

Incorrect Answers

A:

Retain governed immutable fine-tune artifacts and record deployment lineage with tested restore/rollback procedures. This could address fine-tune production endpoint references deleted training artifact, but it does not solve the specified supervised fine-tuning dataset includes conflicting assistant behaviors failure. The matching remediation is: Clean and stratify the supervised fine-tuning data to a coherent supported task format before retraining and evaluating.

B:

Use an appropriate hybrid keyword-plus-vector search strategy with evaluated fusion and optional semantic reranking. This could address pure vector search misses exact regulatory clause codes, but it does not solve the specified supervised fine-tuning dataset includes conflicting assistant behaviors failure. The matching remediation is: Clean and stratify the supervised fine-tuning data to a coherent supported task format before retraining and evaluating.

D:

Halt promotion, review training data and fine-tuning choices, and require safety regression gates on the customized candidate before release. This could address fine-tuned model loses safety boundary after customization, but it does not solve the specified supervised fine-tuning dataset includes conflicting assistant behaviors failure. The matching remediation is: Clean and stratify the supervised fine-tuning data to a coherent supported task format before retraining and evaluating.

Question 598

During a Foundry generative AI quality and observability incident, the following condition occurs: An enterprise support agent appears successful on isolated questions but fails follow-up conversations because the test dataset discards prior chat turns and tool context. The fix must preserve evaluation validity and avoid unnecessary user-data collection. Which targeted Azure AI operations change best addresses the documented problem?

  1. Include appropriate complete conversation histories and tool events in the supported agent evaluation workflow while preserving privacy protections.
  2. Bind relevance evaluation to the authentic user request and assess query rewriting as a separate retrieval-stage metric.
  3. Apply privacy-conscious representative sampling or stratified monitoring that includes error and underrepresented user cohorts.
  4. Add approved prompt-injection and tool-abuse evaluation cases with tool permission gates and block release on unsafe execution.

Explanation

Correct Answer

A:

Include appropriate complete conversation histories and tool events in the supported agent evaluation workflow while preserving privacy protections. This targets multi-turn agent evaluation uses isolated last message only. Multi-turn agents depend on conversation state; evaluating only the final prompt can miss context handling, memory and tool-use errors. Verify the failure through: Examine evaluation records, chat history serialization, tool traces, single-turn versus multi-turn settings and representative failure conversations. Validate success with: The evaluation reveals and measures failures that depend on follow-up context and the selected agent version improves those cases. This matters here because an enterprise support agent appears successful on isolated questions but fails follow-up conversations because the….

Incorrect Answers

B:

Bind relevance evaluation to the authentic user request and assess query rewriting as a separate retrieval-stage metric. This could address relevance evaluator uses wrong intended user query, but it does not solve the specified multi-turn agent evaluation uses isolated last message only failure. The matching remediation is: Include appropriate complete conversation histories and tool events in the supported agent evaluation workflow while preserving privacy protections. This matters here because an enterprise support agent appears successful on isolated questions but fails follow-up conversations because the….

C:

Apply privacy-conscious representative sampling or stratified monitoring that includes error and underrepresented user cohorts. This could address continuous monitoring has no representative sampling coverage, but it does not solve the specified multi-turn agent evaluation uses isolated last message only failure. The matching remediation is: Include appropriate complete conversation histories and tool events in the supported agent evaluation workflow while preserving privacy protections.

D:

Add approved prompt-injection and tool-abuse evaluation cases with tool permission gates and block release on unsafe execution. This could address content safety evaluation omits prompt injection tests, but it does not solve the specified multi-turn agent evaluation uses isolated last message only failure. The matching remediation is: Include appropriate complete conversation histories and tool events in the supported agent evaluation workflow while preserving privacy protections. This matters here because an enterprise support agent appears successful on isolated questions but fails follow-up conversations because the….

Question 599

While deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: An ML team edits training notebooks directly on a shared VM and deploys model versions without recording the corresponding Git commit, preventing a reproducible incident rollback. The change must remain limited to authorized workspaces and avoid unreviewed production impact. Which TWO complementary choices identify the issue and correct it? Select TWO.

  1. Link the correct private DNS zone and approved storage private endpoint to the managed network path before enforcing storage isolation.
  2. Inspect model metadata, job source snapshot, repository history, CI release manifest and available recorded commit identifiers.
  3. Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow.
  4. Inspect Azure deployment operations, Bicep resource references, implicit dependency graph, private endpoint target IDs and pipeline parameter order.
  5. Link the authorized Key Vault or repair supported access to the intended secret store through the correct workspace deployment configuration.

Explanation

Correct Answers

B:

Inspect model metadata, job source snapshot, repository history, CI release manifest and available recorded commit identifiers. This is the most relevant evidence for production training script commit cannot be traced. Source control alone does not provide provenance unless the pipeline records which revision produced the deployed asset. The matching controlled remediation is: Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow. This matters here because while deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: An ML team edits….

C:

Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow. This targets production training script commit cannot be traced. Source control alone does not provide provenance unless the pipeline records which revision produced the deployed asset. Verify the failure through: Inspect model metadata, job source snapshot, repository history, CI release manifest and available recorded commit identifiers. Validate success with: An incident analyst can reproduce the model version's code and dependencies from its recorded source commit. This matters here because while deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: An ML team edits….

Incorrect Answers

A:

Link the correct private DNS zone and approved storage private endpoint to the managed network path before enforcing storage isolation. This is associated with aml private workspace storage access fails from compute, not the essential evidence-and-fix pair for production training script commit cannot be traced. First examine: Inspect model metadata, job source snapshot, repository history, CI release manifest and available recorded commit identifiers. Then correct: Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow.

D:

Inspect Azure deployment operations, Bicep resource references, implicit dependency graph, private endpoint target IDs and pipeline parameter order. This is associated with bicep workspace declaration misses dependent private resources, not the essential evidence-and-fix pair for production training script commit cannot be traced. First examine: Inspect model metadata, job source snapshot, repository history, CI release manifest and available recorded commit identifiers. Then correct: Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow.

E:

Link the authorized Key Vault or repair supported access to the intended secret store through the correct workspace deployment configuration. This is associated with workspace dependencies point to unrelated secrets vault, not the essential evidence-and-fix pair for production training script commit cannot be traced. First examine: Inspect model metadata, job source snapshot, repository history, CI release manifest and available recorded commit identifiers. Then correct: Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow.

Question 600

During a Foundry generative AI quality and observability incident, the following condition occurs: A user-facing agent becomes slow after an LLM deployment change, but the team only records total response time and cannot tell retrieval delay from model generation delay. The fix must preserve evaluation validity and avoid unnecessary user-data collection. Which targeted Azure AI operations change best addresses the documented problem?

  1. Segment continuous evaluation by model/agent deployment version and configure a focused regression alert for the candidate.
  2. Use relevant retrieval context, justified response limits and supported caching after verifying quality on evaluation examples.
  3. Add approved prompt-injection and tool-abuse evaluation cases with tool permission gates and block release on unsafe execution.
  4. Instrument stage-level latency and compare generation and retrieval spans before applying a targeted capacity or retrieval change.

Explanation

Correct Answer

D:

Instrument stage-level latency and compare generation and retrieval spans before applying a targeted capacity or retrieval change. This targets generation latency spike blamed on retrieval without breakdown. Latency budgets should separate agent orchestration, retrieval and generation; total duration alone cannot locate the bottleneck. Verify the failure through: Examine end-to-end request traces, retrieval timing, model time to first token, generation duration, request concurrency and percentiles. Validate success with: The responsible stage is identified and P95 user response time improves without gratuitous changes to healthy components. This matters here because a user-facing agent becomes slow after an LLM deployment change, but the team only records total response time and….

Incorrect Answers

A:

Segment continuous evaluation by model/agent deployment version and configure a focused regression alert for the candidate. This could address foundry continuous evaluation ignores production deployment version, but it does not solve the specified generation latency spike blamed on retrieval without breakdown failure. The matching remediation is: Instrument stage-level latency and compare generation and retrieval spans before applying a targeted capacity or retrieval change.

B:

Use relevant retrieval context, justified response limits and supported caching after verifying quality on evaluation examples. This could address prompt truncation strategy increases generated-token spend, but it does not solve the specified generation latency spike blamed on retrieval without breakdown failure. The matching remediation is: Instrument stage-level latency and compare generation and retrieval spans before applying a targeted capacity or retrieval change.

C:

Add approved prompt-injection and tool-abuse evaluation cases with tool permission gates and block release on unsafe execution. This could address content safety evaluation omits prompt injection tests, but it does not solve the specified generation latency spike blamed on retrieval without breakdown failure. The matching remediation is: Instrument stage-level latency and compare generation and retrieval spans before applying a targeted capacity or retrieval change.

Operationalizing Machine Learning and Generative AI Solutions (AI-300) Practice Tests