Microsoft AI-200 Dumps

Question # 1

For a production AI platform you must achieve end-to-end observability: correlate a user request across an API, a queue, a worker, and model/database calls; alert on latency and error-rate SLOs; and control telemetry cost. Which combination best delivers this?

A.  Autoscale provisioned throughput 
B.  Manual (standard) provisioned throughput sized to the steady load
C.  Serverless 
D.  Strong consistency 

Show Answer

Question # 2

You must evaluate whether a new RAG prompt/version improves answer quality before rolling it out to all users, using production-like traffic and objective metrics such as groundedness and relevance. Which approach is best?

A.  Only test with a handful of manual prompts once 
B.  Deploy the new version to 100% of users immediately and watch for complaints 
C.  Rely solely on unit tests of the code
D.  Run an A/B test routing a portion of traffic to the new version and compare evaluation metrics (groundedness, relevance, latency, cost) before full rollout 

Show Answer

Question # 3

An application intermittently fails to read a rotated secret from Key Vault with '403 Forbidden', but only from certain instances. The identity has the correct role. What should you investigate first?

A. The container image size
B. Key Vault network/firewall settings and private endpoint/VNet configuration for the affected instances, plus RBAC propagation delay
C. The Service Bus lock duration
D. The model deployment quota

Show Answer

Question # 4

A security review requires that no application secrets, connection strings, or keys exist anywhere in code, pipelines, or app settings across your AI solution. Which end-to-end approach best satisfies this?

Rotate keys weekly but keep them in app settings
Put secrets in a private Git repository
Store all secrets as encrypted environment variables in the image
Use managed identities for service-to-service auth (RBAC), Key Vault for the few unavoidable secrets with Key Vault references, and Entra tokens instead of keys

Show Answer

Question # 5

Costs for your Azure OpenAI-backed feature are rising sharply. Investigation shows long prompts with large retrieved context and high max_tokens, plus many repeated identical requests. Which combination best reduces token cost without degrading answer quality?

A. Send the full document corpus in every prompt
B. Trim/re-rank retrieved context to the most relevant chunks, cap max_tokens sensibly, and add a response cache for repeated queries
C. Raise temperature and top_p
D.  Disable retrieval entirely 

Show Answer

Question # 6

Your application receives HTTP 401 from Azure OpenAI after migrating from API keys to managed identity. The managed identity exists and the endpoint is correct. What is the most likely missing step?

A. The model deployment name is wrong
B. The region is not supported
C. The managed identity has not been granted the 'Cognitive Services OpenAI User' role on the resource, orthe client is not requesting an Entra token 
D. The request uses too many tokens

Show Answer

Question # 7

A high-throughput ingestion endpoint must accept millions of telemetry events per hour for later AI analysis, buffering them durably for stream processing. Given the AI-200 messaging options, which service is the best fit for high-volume event streaming ingestion

Azure Service Bus queue with sessions
 Azure App Configuration
Azure Event Hubs (event streaming) as the ingestion buffer
A single Cosmos DB item updated repeatedly

Show Answer

Question # 8

An Event Grid-triggered function must process events exactly in the order they occurred for a given entity, but events are arriving out of order. What is the correct understanding and remedy?

A. Increase the Event Grid retry count to fix ordering
B. Enable duplicate detection to restore order
C. Event Grid guarantees ordering; the function code is buggy
D. Event Grid does not guarantee ordered delivery; use Service Bus sessions (or add sequence numbers and reorder) to enforce per-entity ordering

Show Answer

Question # 9

You are designing idempotent processing for an at-least-once messaging pipeline feeding an AI enrichment step. Which two practices help ensure duplicate deliveries do not cause duplicate side effects? (Choose two.) 

A. Increase the model temperature for variety
B. Make write operations idempotent using deterministic keys/upserts
C. Use ReceiveAndDelete so messages are never redelivered
D. Track processed message IDs (for example, in Cosmos DB or Redis) and skip already-seen IDs

Show Answer

Question # 10

An Azure Function on the Consumption plan experiences cold-start latency that violates a strict p99 requirement for a customer-facing AI endpoint, and it must also reach resources over a private VNet. Which change best addresses both concerns?

A. Increase function timeout
B. Move to the Premium plan (pre-warmed instances and VNet integration)
C. Keep Consumption plan and add more logging
D. Switch to a Timer trigger

Show Answer

Question # 11

A solution must fan a single 'DocumentIngested' event out to three independent consumers (indexing, notification, and audit), each able to fail and retry independently without affecting the others. Which design is most appropriate?

One Service Bus queue read by all three consumers competing for messages
Event Grid topic with three subscriptions, or a Service Bus topic with three subscriptions, so each consumer gets its own copy
A single function that calls all three synchronously
A Redis list shared by all consumers

Show Answer

Question # 12

A Service Bus-triggered function's messages keep reappearing and are eventually dead-lettered with reason 'MaxDeliveryCountExceeded', even though processing seems to complete. Logs show processing sometimes takes longer than the lock duration. What is the best fix?

A. Disable dead-lettering
B. Renew the message lock during long processing (or increase lock duration) so the lock does not expire before completion
C. Lower MaxDeliveryCount to 1
D.  Switch to Event Grid 

Show Answer

Question # 13

An order-processing pipeline must guarantee that duplicate submissions (same OrderId sent twice within 10 minutes) are not processed twice, and that messages for the same OrderId are handled in order. Which Service Bus configuration meets both needs?

A. Use ReceiveAndDelete mode with prefetch
B. Enable partitioning and set a large max size
C. Enable duplicate detection with a 10-minute window and use sessions keyed by OrderId
D. Use Event Grid with retries

Show Answer

Question # 14

A RAG pipeline returns answers that are fluent but sometimes unsupported by the retrieved context (hallucinations). Retrieval quality metrics show low groundedness. Which combination of changes most directly improves groundedness?

Switch the vector store from Cosmos DB to Redis
Improve chunking/retrieval relevance, add hybrid + re-ranking, and instruct the model to answer only from provided context (and cite sources)
Remove the system prompt entirely
Increase temperature and max_tokens

Show Answer

Question # 15

After adding an IVFFlat index to a growing pgvector table, recall (result quality) has degraded even though queries are faster. What is the most likely cause and remedy?

A. The embeddings must be stored as JSON
B. The IVFFlat 'lists'/'probes' settings are not tuned for the current data size; increase probes (or rebuild withmore lists) to improve recall 
C. The table needs a B-tree index instead
D. pgvector does not support similarity search

Show Answer

Question # 16

A financial RAG assistant must never mix data between customers, must audit every retrieval, and must minimize the risk of one customer's vectors surfacing for another. Using Cosmos DB for NoSQL, which design most reduces cross-tenant leakage risk?

A. A single global index with client-side filtering after retrieval
B. Store all tenants in one item to simplify queries
C. One shared container with customerId only used as a query filter and no partitioning by tenant
D. Partition by tenant (customerId) and always scope vector queries to that partition key, with retrieval logging

Show Answer

Question # 17

You must lower the cost and latency of pgvector similarity search on a large, growing table while keeping recall acceptable. Which two actions are appropriate? (Choose two.)

A. Build an HNSW index and tune its parameters (for example, m and ef_search)
B. Force sequential scans for every query
C. Remove the vector column and store embeddings in TEXT
D. Right-size compute/memory so the index and working set fit in RAM

Show Answer

Question # 18

A retrieval service using Azure Managed Redis must implement a semantic cache: if a new question is highly similar to a previously answered one, return the cached answer; otherwise call the model. Which Redis capability enables the 'highly similar' lookup?

A. Simple string GET on the exact question text
B. Redis Streams consumer groups
C. Redis vector indexing/similarity search over stored question embeddings with a similarity threshold
D. Redis pub/sub channels

Show Answer

Question # 19

An enterprise RAG system must refresh its vector index within seconds whenever source documents in Cosmos DB change, without a nightly batch job. Which architecture is best? 

A.  A change feed processor that, on each changed item, re-chunks/re-embeds only the affected document and upserts its vectors 
B.  A Timer function that re-embeds all documents every night 
B.  A Timer function that re-embeds all documents every night 
D.  Manual re-indexing triggered by an operator 

Show Answer

Question # 20

A Cosmos DB for NoSQL container shows one physical partition consuming most RU/s and frequent 429s, while others are nearly idle. Point reads by 'id' are fine, but throughput cannot be raised cost-effectively. What is the best remediation?

A.  Redesign the partition key to one with high cardinality and even access (for example, a synthetic/composite key) and migrate data 
B.  Switch to eventual consistency
C.  Increase provisioned RU/s until 429s stop
D.  Enable the analytical store

Show Answer