Azure AI-200 Cheat Sheet

Containerized Solutions on Azure

20-25% of exam
Container Registry App Service Containers Container Apps Scaling and KEDA AKS Essentials Compute Picker

AI Solutions with Data Services

25-30% of exam
Cosmos DB for NoSQL Cosmos Vector Search pgvector on PostgreSQL Azure Managed Redis Data Store Picker Vector Index Limits

Connect and Consume Azure Services

20-25% of exam
Service Bus Event Grid Azure Functions Messaging Picker Trigger vs Binding

Secure, Monitor, Troubleshoot

20-25% of exam
Key Vault + App Config Identity and Access Tracing and App Insights KQL Query Basics Troubleshoot Picker

Quick Facts

Exam AI-200
Credential Azure AI Cloud Developer Associate
Time 120 min
Pass 700 of 1000
Score type Scaled, not 70%
Questions Not published by Microsoft
Level Associate
Language English only
Fee Varies by country
Renewal Free yearly assessment
Retake 24 hours, then 14 days
Blueprint Apr 15 2026

Container Apps Defaults CO

No rule: HTTP, 0 to 10
10 concurrent requests Poll every 30 seconds Cooldown 300 seconds No ingress? minReplicas 1

Container Apps vs AKS CO

Container Apps

Serverless, no cluster
KEDA built in
Scales to zero

AKS

Full Kubernetes control
Manifests and kubectl
You manage nodes
Managed vs full control

Compute Picker CO

Need full cluster control → AKS ( Manifests, kubectl )
Containers without Kubernetes → Container Apps ( Serverless )
Single web container → App Service ( PaaS )
Event-driven code only → Azure Functions ( Serverless )
Run to completion → Container Apps job ( Batch )
Build without local Docker → az acr build ( ACR Tasks )
Blue-green or canary → Multiple revision mode ( Traffic split )
Scale on queue depth → KEDA custom rule ( Service Bus )

Container Registry CO

ACR Private image registry
az acr build Builds image in cloud No Docker
ACR Tasks Automated cloud builds
Base image trigger Rebuilds on patch
AcrPull Role that pulls images
Admin account Shared password, keep disabled
Premium tier Geo-replication, private endpoints
Image digest Immutable content hash
Image tag Mutable pointer

Single vs Multiple Revision CO

Single

Default mode
New revision takes all
No traffic split

Multiple

Several revisions active
Weighted traffic split
Blue-green and canary
Splitting traffic needs multiple

App Service Containers CO

WEBSITES_PORT Port the container listens
App settings Injected as environment variables
Key Vault reference Secret resolved at startup
Managed identity pull Keyless registry access
Deployment slot Stage, then swap
Slot setting Stays with its slot
Always On Stops idle unload
Health check Removes unhealthy instances

Template vs Configuration Scope CO

Template

Image, env, resources
Scale rules
Creates new revision

Configuration

Ingress, secrets, registry
Revision mode
Applies to all
New revision vs global

Container Apps CO

Environment Network and logging boundary
Revision Immutable app snapshot
Template change Creates a new revision
Configuration change Applies to all revisions
Ingress External or internal
Dapr sidecar Pub/sub, state, bindings
Jobs Run to completion
Secrets Referenced by environment variables
Workload profile Dedicated or consumption compute

Scaling and KEDA CO

KEDA Event-driven autoscaler
Default rule HTTP, 0-10 replicas Default
concurrentRequests Default 10 per replica
minReplicas 0 Scales to zero
No ingress risk Set minReplicas to 1
Custom scaler Service Bus, Kafka, Redis
Polling interval 30 seconds
Cooldown period 300 seconds
Scale formula ceil(current / target)

AKS Essentials CO

kubectl apply -f Deploys a manifest
Deployment Manages replica pods
Service Stable network endpoint
Ingress HTTP routing layer
ConfigMap Non-secret configuration
Secret Base64 encoded, not encrypted
kubectl logs Container stdout output
kubectl describe Events and failure reasons
Workload identity Pod-level Entra token
az aks update --attach-acr Grants AcrPull to cluster

Domain Weights DA

Data leads; other three tie
Containers: 20-25% Data services: 25-30% Connect: 20-25% Secure: 20-25%

Cosmos Vector vs pgvector DA

Cosmos NoSQL

JSON with embeddings
VectorDistance function
diskANN at scale

PostgreSQL

Relational joins and filters
<-> and <=> operators
HNSW or IVFFlat
Documents vs relational

Data Store Picker DA

Relational plus vectors → PostgreSQL + pgvector ( SQL joins )
Schema-free JSON documents → Cosmos DB NoSQL ( Global scale )
Sub-millisecond lookups → Azure Managed Redis ( In-memory )
Millions of vectors → diskANN index ( Cosmos DB )
Exact recall, small set → flat index ( 505 dimensions )
Under 50,000 vectors → quantizedFlat index ( Lower RU )
Best recall in Postgres → HNSW index ( Slower build )
Faster build in Postgres → IVFFlat index ( Tune probes )
React to item changes → Change feed processor ( Cosmos DB )

Cosmos DB for NoSQL DA

RU/s Throughput currency
Partition key Spreads logical partitions
Autoscale Ranges max/10 to max
Session Default consistency level Default
Strong Highest RU cost
Indexing policy Include or exclude paths
TTL Automatic item expiry
Change feed Ordered record of changes
_ts Modification timestamp
_etag Optimistic concurrency token

pgvector Operators DA

<-> L2, <=> cosine, <#> inner
<-> Euclidean distance <=> cosine distance <#> negative inner product

HNSW vs IVFFlat DA

HNSW

Graph based
Best recall
Slower, larger build

IVFFlat

Cluster lists
Faster build
Needs representative data
Recall vs build cost

Cosmos Vector Search DA

Container vector policy path, dataType, dimensions, distance
VectorDistance() Similarity system function
flat Exact search, 505 dimensions
quantizedFlat Compressed, up to 4096
diskANN Approximate, above 50,000 vectors
distanceFunction cosine, dotproduct, euclidean
TOP N Always cap vector queries
1,000 vector floor Needed by quantized indexes
Default dimensions 1536

Vector Index Limits DA

flat 505, quantizedFlat 4096, diskANN 4096
flat: exact, small sets quantizedFlat: under 50,000 diskANN: over 50,000 Quantized needs 1,000 vectors

flat vs diskANN DA

flat

Brute force scan
100% recall
Max 505 dimensions

diskANN

Approximate search
Max 4096 dimensions
Needs 1,000 vectors
Exact vs scalable

pgvector on PostgreSQL DA

azure.extensions Allowlist before creating
CREATE EXTENSION vector Extension name is vector
vector(n) Column type, n dimensions
<-> Euclidean L2 distance
<=> Cosine distance
<#> Negative inner product
HNSW Graph index, best recall
IVFFlat List index, faster build
hnsw.ef_search Recall versus latency knob
ivfflat.probes Lists scanned per query
Metadata filter WHERE clause narrows RAG

Consistency Ladder DA

Strong > Bounded > Session > Prefix > Eventual
Session is default Strong costs most RUs Eventual is cheapest

Azure Managed Redis DA

Cache-aside Load on cache miss
EXPIRE Time-based invalidation
Eviction policy What drops when full
FT.CREATE Builds a search index
Vector index types FLAT or HNSW
Tiers Memory, Balanced, Compute, Flash
Flash Optimized No search or vector
Entra ID auth Keyless Redis access
Redis version 7.4.x
Semantic cache Vector-matched response reuse

Event Grid Retry CO

10s, 30s, 1m, 5m, then hours
TTL default 1440 minutes 30 attempts default Dead-letter needs storage

Service Bus vs Event Grid CO

Service Bus

Carries commands and work
Ordering, sessions, DLQ
Consumer pulls

Event Grid

Announces facts
Pushes with filters
At-least-once, unordered
Work queue vs notification

Messaging Picker CO

Ordered transactional work → Service Bus queue ( Sessions )
Fan-out with filters → Service Bus topic ( Subscriptions )
Reactive discrete events → Event Grid ( CloudEvents )
Poison message handling → Dead-letter queue ( MaxDeliveryCount )
Serverless HTTP API → Functions HTTP trigger ( Binding )
Multi-step workflow state → Durable Functions ( Orchestrator )
Delay a message → Scheduled enqueue time ( Service Bus )

Service Bus CO

Queue Point-to-point work
Topic Publish to many subscriptions
Subscription filter SQL or correlation rule
DLQ Holds undeliverable messages
MaxDeliveryCount Moves message to DLQ
Peek-lock Receive, process, then complete
Sessions FIFO by session ID
Duplicate detection Ignores repeated MessageId
Standard size 256 KB message
Premium size Up to 100 MB

Function Timeouts CO

Consumption 5/10; Flex and Premium 30
Consumption: 5 then 10 Flex: 30, then unbounded HTTP capped 230 seconds

Event Grid vs Event Hubs CO

Event Grid

Discrete events
Per-event delivery
Reactive handlers

Event Hubs

High-volume streams
Partitions and offsets
Telemetry ingestion
Discrete vs stream

Event Grid CO

CloudEvents 1.0 Standard event schema
Subject filter Prefix or suffix match
Advanced filter Match on data fields
At-least-once Duplicates possible, order not guaranteed
Retry policy Exponential backoff, best effort
Event TTL Default 1440 minutes
Max attempts Default 30 tries
Dead-letter Storage container required
Success codes 200 through 204 only
Never retried 400, 403, 413

Trigger vs Binding CO

Trigger

Starts execution
Exactly one
Carries payload

Binding

Input or output
Zero or many
Cannot start function
Starts vs connects

Azure Functions CO

Trigger Starts the function
Input binding Reads data declaratively
Output binding Writes data declaratively
host.json App-wide runtime settings
local.settings.json Local development only
Flex Consumption Current serverless plan Default
Consumption timeout 5 default, 10 max
Flex and Premium 30 default, unbounded max
HTTP response cap 230 seconds regardless
Durable Functions Stateful orchestration patterns

Keyless Access Recipe SE

Identity, role, scope, credential
Enable managed identity Assign least-privilege role Scope to the resource Use DefaultAzureCredential

Key Vault vs App Configuration SE

Key Vault

Secrets, keys, certificates
Audited and rotated
Near-expiry events

App Configuration

Non-secret settings
Labels and snapshots
Feature flags
Secrets vs settings

Troubleshoot Picker SE

Slow end-to-end request → Distributed trace ( App Insights )
Which service failed → Application Map ( Dependencies )
Search logs at scale → KQL query ( Log Analytics )
Container keeps restarting → Container console logs ( System logs )
Secret will not resolve → Check identity role ( RBAC )
Traffic right now → Live Metrics ( Streaming )
Traces not correlating → traceparent header ( W3C context )

Key Vault + App Config SE

Secret Password, token, connection string
Soft delete Recoverable, cannot be disabled
Purge protection Blocks early permanent delete
RBAC data plane Preferred over access policies
Rotation Near-expiry event triggers function
Key Vault reference App setting resolves secret
App Configuration Central non-secret settings
Label Per-environment key variant
Sentinel key Signals refresh of all
Feature flag Runtime on/off toggle

System vs User Assigned SE

System-assigned

One resource only
Deleted with resource
Simplest single app

User-assigned

Standalone resource
Shared by many
Survives redeploys
Lifecycle tied vs independent

Identity and Access SE

Managed identity Azure-managed, no secrets
System-assigned Tied to one resource
User-assigned Shared across resources
DefaultAzureCredential Chained credential lookup
Workload identity Federated token for pods
OIDC federation GitHub Actions without secrets
AcrPull Pulls container images
Key Vault Secrets User Reads secret values
Cosmos DB Built-in Data Reader Data plane read role
Private endpoint Private IP, no internet

Tracing and App Insights SE

OpenTelemetry Vendor-neutral telemetry standard
Azure Monitor Distro Microsoft OpenTelemetry package
Connection string Routes telemetry to resource
traceparent W3C trace context header
Span One unit of work
Cloud role name Names service on map
Application Map Service dependency topology
requests Incoming operation telemetry
dependencies Outbound call telemetry
exceptions Captured error telemetry
Live Metrics Real-time streaming view

KQL Query Basics SE

where Filters rows
project Selects columns
extend Adds computed column
summarize Aggregates by group
bin() Buckets into time intervals
ago() Relative time filter
top Highest N rows
join Combines two tables
union Stacks multiple tables
render Charts the result
let Names a variable

Common Traps

Scaled score

700 of 1000 scaled ≠ Not 70% correct

Identity vs role

Managed identity authenticates ≠ RBAC role authorizes

Work vs notification

Service Bus carries work ≠ Event Grid announces facts

Policy vs index

Vector policy defines shape ≠ Vector index speeds search

Secret vs setting

Key Vault holds secrets ≠ App Config holds settings

Revision vs configuration

Template edits create revisions ≠ Configuration edits do not

Zero replica dead end

No ingress, no rule ≠ App cannot wake up

Distance vs similarity

Lower distance is closer ≠ Cosine distance, not similarity

Flash tier limit

Flash Optimized uses NVMe ≠ No search or vector

Extension name

Docs say pgvector ≠ SQL says CREATE EXTENSION vector

Last Minute

1. Pass = 700 scaled, not 70%
2. 120 minutes, English only
3. Data services is biggest domain
4. Container Apps default: 0-10 replicas
5. No ingress? Set minReplicas 1
6. Multiple revision mode splits traffic
7. az acr build needs no Docker
8. flat 505 dims; diskANN 4096
9. VectorDistance queries need TOP N
10. pgvector: <-> L2, <=> cosine
11. HNSW = recall; IVFFlat = speed
12. Service Bus = work; Event Grid = events
13. Event Grid TTL 1440 minutes
14. DLQ triggered by MaxDeliveryCount
15. Consumption timeout 5, max 10
16. Key Vault secrets; App Config settings
17. Managed identity first, then RBAC
18. traceparent carries W3C trace context
19. KQL: where, summarize, project, render
20. Flash Optimized has no vector
21. Renew free every 12 months

Microsoft Azure AI Cloud Developer (AI-200)

Practice Questions 218 questions Study Guide 35 topics Cheat Sheet 157 facts Flashcards 50 cards Videos 1 video Blogs 1 blog Free CE Podcast