AI-200 — Microsoft Azure AI Cloud Developer Associate Quick Reference

Last revised: June 18, 2026

Compact AI-200 quick reference for Azure AI service selection, RAG patterns, security, deployment, and troubleshooting.

How to Use This Quick Reference

This independent Quick Reference is for candidates preparing for the Microsoft Azure AI Cloud Developer Associate (AI-200) exam. Use it to review decision points, service selection, implementation patterns, and common traps for Azure AI development scenarios.

Focus less on memorizing product names and more on answering: Which Azure AI service fits the requirement, how is it secured, how is it deployed, and how do you evaluate and troubleshoot it?

High-Yield Exam Map

Area	What to know for AI-200-style scenarios
Azure AI service selection	Choose between Azure OpenAI, Azure AI Search, Azure AI Document Intelligence, Azure AI Language, Azure AI Vision, Azure AI Speech, Translator, Content Safety, Azure Machine Learning, and app hosting services.
Generative AI development	Chat completions, embeddings, model deployments, prompt structure, tools/function calling, token management, response grounding, and evaluation.
Retrieval-augmented generation	Chunking, embeddings, vector indexes, hybrid search, semantic ranking, citations, access filtering, freshness, and hallucination reduction.
Knowledge mining	Azure AI Search indexes, indexers, data sources, skillsets, enrichment pipelines, custom skills, and semantic/vector search.
Natural language, speech, vision, documents	Select prebuilt vs custom models; distinguish OCR, form extraction, image analysis, transcription, translation, classification, and entity extraction.
Security and governance	Microsoft Entra ID, managed identities, keys, Key Vault, RBAC, private networking, content filters, responsible AI controls, logging, and data protection.
Deployment and operations	App Service, Azure Functions, Container Apps, AKS, API Management, queues, monitoring, retries, throttling, testing, and CI/CD.

Azure AI Service Selection Matrix

Requirement	Prefer	Why	Watch for
Build a chat, summarization, reasoning, or code-assist feature	Azure OpenAI Service or model deployments through Azure AI development tooling	Managed access to large language models with Azure security and deployment controls	The model name is not enough; apps call a deployment. Region and model availability matter.
Build, test, evaluate, and manage generative AI apps	Azure AI Foundry tooling	Project-based development, prompt workflows, evaluations, deployments, and model catalog workflows	Do not confuse design-time project tooling with the runtime app architecture.
Ground an LLM on enterprise documents	Azure AI Search + embeddings + Azure OpenAI	Supports keyword, vector, hybrid, semantic ranking, metadata filters, and citations	Retrieval does not guarantee correctness; still evaluate groundedness and safety.
Search structured and unstructured enterprise content	Azure AI Search	Indexes documents, supports filters, scoring, semantic ranking, vector search, and enrichment	Index schema, analyzer choice, vector dimensions, and metadata fields are exam-relevant.
Extract fields from invoices, receipts, IDs, tax forms, or custom forms	Azure AI Document Intelligence	Prebuilt and custom document extraction models	Use Document Intelligence for field extraction, not generic OCR-only scenarios.
OCR text from images or simple documents	Azure AI Vision Read/OCR or Document Intelligence	Vision handles image OCR; Document Intelligence handles document-centric extraction	If the scenario needs key-value pairs/tables/forms, choose Document Intelligence.
Analyze images for captions, tags, objects, or visual features	Azure AI Vision	Prebuilt image analysis capabilities	Do not choose Document Intelligence for general image tagging.
Classify images with custom labels	Custom Vision / custom image model workflow	Train image classification or object detection from labeled images	Use only when prebuilt Vision features are insufficient.
Detect language, sentiment, key phrases, entities, or PII	Azure AI Language	Prebuilt NLP APIs	Use custom Language models when domain-specific labels, intents, or entities are needed.
Build intent recognition for a chatbot	Conversational Language Understanding	Maps user utterances to intents and entities	CLU identifies intent; it does not automatically complete business workflows.
Create FAQ-style question answering over curated content	Custom question answering	Best for controlled knowledge bases and FAQ-style responses	For broad document retrieval plus generation, prefer RAG with Azure AI Search and an LLM.
Translate text between languages	Translator	Purpose-built machine translation	Do not use speech translation unless audio is involved.
Transcribe or synthesize speech	Azure AI Speech	Speech-to-text, text-to-speech, speech translation, custom speech scenarios	Batch vs real-time and custom model requirements are common decision points.
Detect harmful, unsafe, or policy-violating content	Azure AI Content Safety plus model content filters	Safety classification for text/images and layered protection for generative AI apps	Safety filters reduce risk; they are not a substitute for app authorization or validation.
Train, register, deploy, and monitor custom ML models	Azure Machine Learning	Full ML lifecycle for custom models, pipelines, endpoints, and MLOps	Do not choose Azure ML when a prebuilt Azure AI service satisfies the requirement.
Expose AI functionality through an API	App Service, Azure Functions, Container Apps, AKS + API Management	Hosts app logic and protects/standardizes APIs	The AI service is not usually the entire application boundary.
Trigger AI processing from uploaded files/events	Event Grid, Service Bus, Storage Queue, Azure Functions	Event-driven ingestion and asynchronous processing	Use queues for buffering, retries, and decoupling long-running AI tasks.

Core Azure AI Terms

Term	Meaning	Exam distinction
Azure AI services resource	Azure resource used to access one or more cognitive services APIs	Multi-service resources simplify management but not every scenario uses one shared endpoint.
Azure OpenAI resource	Azure resource for deploying and calling OpenAI models through Azure	You deploy a model before an app can call it.
Model	The base AI capability, such as a chat model or embedding model	Model availability is not the same as deployment availability.
Deployment	Named runtime instance of a model in an Azure OpenAI resource	In many SDK calls, the `model` parameter is the deployment name.
Endpoint	Network address used by applications to call a service	May be public, restricted by firewall, or private through Private Link.
Key	Shared secret for API access	Simpler but weaker operational model than managed identity. Store in Key Vault when used.
Managed identity	Microsoft Entra identity assigned to an Azure workload	Preferred for Azure-hosted apps calling Azure services without secrets.
RBAC	Role-based access control through Microsoft Entra ID	Separate management-plane permissions from data-plane permissions.
Index	Searchable structure in Azure AI Search	Requires a schema, fields, analyzers, and optionally vector fields.
Indexer	Crawler that loads data from a supported source into an index	Runs on schedule or demand; does not run at query time.
Skillset	Enrichment pipeline for Azure AI Search indexing	Applies OCR, extraction, translation, custom skills, or projections during indexing.
Embedding	Numeric vector representation of text/images	Query and document embeddings must be generated with compatible models and dimensions.
Chunk	Segment of a document indexed for retrieval	Bad chunking causes weak grounding even with a strong model.
Semantic ranking	Language-aware ranking layer in Azure AI Search	Often combined with keyword/vector retrieval for better relevance.
Content filter	Safety control applied to model inputs/outputs	Not an authorization system and not a full business policy engine.

Generative AI Development Reference

Chat Completion Anatomy

Component	Purpose	Common trap
System instruction	Sets assistant behavior, constraints, tone, and task rules	It is not a security boundary. Always validate inputs, outputs, and tool calls.
User message	End-user request	User content may contain prompt injection attempts.
Assistant message	Prior model response	Long histories consume tokens and may preserve bad context.
Tool/function definition	Describes callable app functions	The model suggests calls; your code authorizes and executes them.
Retrieved context	External data inserted into the prompt	Must be relevant, access-controlled, and cited when required.
Response format	Controls structured output, such as JSON	Validate schema after generation; do not assume perfect formatting.
Temperature/top-p	Controls randomness	Lower values usually suit extraction, classification, and deterministic business tasks.
Max tokens	Caps response length	Too low truncates answers; too high can increase latency and cost.

Azure OpenAI SDK Pattern

Use Microsoft Entra ID or managed identity where possible for production workloads. API keys are common in simple examples but should be protected.

from azure.identity import DefaultAzureCredential, get_bearer_token_provider
from openai import AzureOpenAI

token_provider = get_bearer_token_provider(
    DefaultAzureCredential(),
    "https://cognitiveservices.azure.com/.default"
)

client = AzureOpenAI(
    azure_endpoint="https://<resource-name>.openai.azure.com/",
    azure_ad_token_provider=token_provider,
    api_version="<api-version>"
)

response = client.chat.completions.create(
    model="<deployment-name>",
    messages=[
        {"role": "system", "content": "Answer using the provided policy excerpt only."},
        {"role": "user", "content": "What is the refund window?"}
    ],
    temperature=0.2
)

print(response.choices[0].message.content)

High-yield detail: in Azure OpenAI calls, model="<deployment-name>" commonly refers to the Azure deployment name, not just the base model family.

Model Choice Decision Points

Need	Prefer	Notes
General chat, summarization, reasoning, extraction	Chat/completion model	Choose based on required quality, latency, cost, context length, and region availability.
Enterprise RAG	Chat model + embedding model + Azure AI Search	The chat model generates; the embedding model retrieves.
Similarity search	Embedding model	Store embeddings in a vector index; do not ask embeddings to generate answers.
Deterministic classification/extraction	Lower temperature, schema validation, possibly Azure AI Language or Document Intelligence	For standard NLP/document tasks, prebuilt services may be more reliable and simpler.
Multimodal reasoning	Model/service that supports the required input type	Verify whether the scenario needs image, text, audio, or document-native processing.
High-volume automation	Smaller/faster model where acceptable, caching, batching, queueing	Avoid using the largest model by default.
Regulated or sensitive workflow	Private networking, managed identity, logging strategy, human review, content safety	Security and governance often decide the architecture.

Retrieval-Augmented Generation Reference

RAG Flow

    flowchart LR
	    A[Source documents] --> B[Extract text and metadata]
	    B --> C[Chunk documents]
	    C --> D[Generate embeddings]
	    D --> E[Index in Azure AI Search]
	    U[User question] --> V[Embed query]
	    V --> W[Vector / hybrid retrieval]
	    W --> X[Rank, filter, and trim]
	    X --> Y[Prompt with retrieved context]
	    Y --> Z[Generate grounded answer with citations]
	    Z --> Q[Evaluate, log, and monitor]

RAG Design Matrix

Design choice	Use when	Exam cues	Traps
Keyword search	Exact terms, IDs, names, codes, or structured phrases matter	“Find documents containing…”	Poor semantic recall for paraphrased questions.
Vector search	Users ask semantically similar but differently worded questions	“Natural language questions over documents”	Vector dimensions must match the embedding model.
Hybrid search	Need both exact matching and semantic recall	“Best relevance over enterprise content”	Requires tuning scoring, filters, and ranking.
Semantic ranking	Need improved natural-language relevance and captions/answers	“Improve result quality without retraining”	It ranks retrieved candidates; it does not replace indexing.
Metadata filtering	Need access control, departments, dates, regions, document types	“Only show documents user can access”	Filter fields must exist and be populated in the index.
Security trimming	Results must respect user permissions	“User-specific document access”	Do not rely on the LLM to hide unauthorized text after retrieval.
Chunk overlap	Concepts span boundaries between chunks	“Answers miss context at page breaks”	Too much overlap increases index size and duplicate retrieval.
Citations	Users need traceability	“Answer with sources”	Citations require source metadata captured during ingestion.
Freshness	Data changes often	“New documents must appear quickly”	Scheduled indexers may not meet near-real-time needs; consider push/event ingestion.
Human review	High-impact or risky outputs	“Approval required before action”	Content filters alone may be insufficient.

Minimal Search Index Fields for RAG

Field	Purpose	Search configuration
`id`	Stable unique key	Key field
`content`	Chunk text passed to the model	Searchable
`contentVector`	Embedding for vector search	Vector field with matching dimensions
`title`	Human-friendly source label	Searchable/filterable as needed
`sourceUri`	Citation link or storage reference	Retrievable
`pageNumber` / `section`	Citation precision	Filterable/retrievable
`lastModified`	Freshness filtering/sorting	Filterable/sortable
`acl` / `groups`	Security trimming	Filterable
`documentType`	Filter by policy, manual, contract, etc.	Filterable/facetable

Vector/Hybrid Search SDK Shape

from azure.search.documents import SearchClient
from azure.search.documents.models import VectorizedQuery
from azure.identity import DefaultAzureCredential

search_client = SearchClient(
    endpoint="https://<search-service>.search.windows.net",
    index_name="<index-name>",
    credential=DefaultAzureCredential()
)

vector_query = VectorizedQuery(
    vector=query_embedding,
    k_nearest_neighbors=5,
    fields="contentVector"
)

results = search_client.search(
    search_text="refund policy for annual subscriptions",
    vector_queries=[vector_query],
    filter="documentType eq 'policy'",
    select=["title", "content", "sourceUri", "pageNumber"],
    top=5
)

Use this pattern to remember the separation between query embedding, vector retrieval, metadata filtering, and prompt construction.

Azure AI Search and Knowledge Mining

Component	Role	When to use	Common issue
Data source	Connection to supported content store	Pull data from Azure Storage, databases, or other supported sources	Permissions and private networking can block indexers.
Indexer	Moves data into an index	Scheduled or on-demand indexing	It does not continuously reflect changes unless scheduled or triggered.
Skillset	Enriches content during indexing	OCR, entity extraction, key phrases, translation, custom enrichment	Skills run at ingestion time, not at query time.
Custom Web API skill	Calls your custom enrichment logic	Domain-specific extraction, normalization, classification	Must handle scaling, failures, and expected schema.
Index projection	Maps enriched content into target index structures	Parent-child or chunked indexing patterns	Incorrect mapping leads to missing fields.
Analyzer	Tokenization and text processing	Language-specific search behavior, stemming, tokenization	Analyzer choice affects matching and cannot always be casually changed later.
Synonym map	Expands equivalent terms	Industry acronyms, product aliases	Synonyms help keyword search but do not replace semantic/vector search.
Semantic configuration	Defines prioritized fields for semantic ranking	Better captions/reranking for natural-language queries	Needs meaningful title/content fields.
Vector profile/configuration	Enables vector search	Embedding-based retrieval	Embedding dimensions and vector field config must align.

Knowledge Mining vs RAG

Scenario	Better answer
“Extract entities and key phrases from documents into a searchable index”	Azure AI Search skillset with enrichment
“Ask natural-language questions and generate answers from indexed documents”	RAG using Azure AI Search plus Azure OpenAI
“Search documents with filters, facets, and relevance scoring”	Azure AI Search
“Summarize search results into a conversational response”	Azure AI Search retrieval followed by generative model response
“Apply OCR before indexing scanned PDFs”	Azure AI Search skillset with OCR, or Document Intelligence depending on extraction needs

Language, Speech, Vision, and Document Services

Azure AI Language

Requirement	Choose	Notes
Sentiment and opinion mining	Sentiment analysis	Identifies positive/negative/neutral sentiment and opinions where supported.
Extract names, places, organizations, dates	Named entity recognition	Use custom NER for domain-specific entities.
Detect sensitive personal data	PII detection	Combine with app policy for redaction, storage, and auditing.
Extract important terms	Key phrase extraction	Useful for tagging and indexing.
Identify language	Language detection	Often used before translation or language-specific processing.
Classify text into custom categories	Custom text classification	Requires labeled examples and training/evaluation.
Extract domain-specific entities	Custom named entity recognition	Use when prebuilt NER misses business-specific labels.
Detect user intent and entities in conversations	Conversational Language Understanding	Good for bot commands and routing.
FAQ-style answers from curated sources	Custom question answering	Best for controlled knowledge base scenarios.

Azure AI Speech and Translator

Requirement	Choose	Key distinction
Convert microphone or audio files to text	Speech-to-text	Real-time vs batch transcription matters.
Convert text to spoken audio	Text-to-speech	Voice, language, and style requirements drive selection.
Translate text	Translator	Text input/output.
Translate spoken audio	Speech translation	Audio input with translation output.
Improve recognition for domain vocabulary	Custom Speech	Use when baseline transcription struggles with accents, terms, or environment.
Build voice-enabled app	Speech SDK + app host	The SDK handles audio interaction; your app handles business logic.

Azure AI Vision and Document Intelligence

Requirement	Choose	Avoid this mistake
Read text from an image	Azure AI Vision OCR/Read	Do not build a custom model for basic OCR.
Extract fields from forms	Azure AI Document Intelligence	OCR alone does not produce structured fields reliably.
Extract tables from documents	Document Intelligence	Tables require document-aware layout extraction.
Use prebuilt invoice/receipt/ID extraction	Document Intelligence prebuilt model	Do not train custom if a prebuilt model satisfies the form type.
Extract from a custom business form	Document Intelligence custom model	Needs representative labeled samples and evaluation.
Classify document types before extraction	Document classifier / routing pattern	Route to the right extraction model.
Generate image tags/captions	Azure AI Vision image analysis	Document Intelligence is document-centric, not image-scene analysis.
Detect custom objects in images	Custom Vision or custom vision model workflow	Requires labeled images and model training.

Prompting, Tools, and Agentic Patterns

Pattern	Use when	Implementation reminder
Direct prompt	Simple transformation, drafting, summarization, or classification	Keep instructions explicit and constrain output format.
Few-shot prompt	Need consistent style or labels	Include representative examples; avoid excessive token use.
RAG prompt	Need answers grounded in private/current data	Retrieve first, then generate using context and citation instructions.
Tool/function calling	The model needs live data or actions	Validate arguments, authorize user, execute tool in app code, then return result to model.
Planner/agent loop	Multi-step tasks with tool use	Add iteration limits, logging, timeout, safety checks, and human approval for risky actions.
Structured output	Downstream system expects JSON or schema	Validate and retry/repair; never blindly trust generated JSON.
Prompt template	Reusable prompt with variables	Treat retrieved/user content as data, not instructions.
Guardrail prompt	Reduce unsafe behavior	Helpful but insufficient without content filters, authorization, and validation.

Tool Calling Control Points

Step	Control
Tool definition	Expose only required functions and arguments.
User request	Authenticate user and check authorization before tool execution.
Model-proposed tool call	Validate name, arguments, types, ranges, and policy constraints.
Tool execution	Use least-privilege identity and handle timeouts/retries.
Tool result	Remove secrets and unnecessary data before sending back to the model.
Final response	Check safety, correctness, citation, and formatting requirements.

Security, Identity, and Governance

Authentication and Authorization Choices

Option	Best use	Exam distinction
API key	Simple local testing or services that require key-based access	Store in Key Vault; rotate; avoid embedding in code or client apps.
Microsoft Entra ID	Enterprise authentication and RBAC	Preferred for production when supported.
Managed identity	Azure-hosted app calling Azure services	Avoids secrets; assign least-privilege roles.
Service principal	CI/CD or non-Azure workload automation	Protect credentials; scope permissions tightly.
SAS token	Limited delegated access to storage objects	Time-bound and permission-scoped; not an identity replacement.
Key Vault	Secret, key, and certificate management	App identity needs permission to retrieve secrets.

RBAC and Access Boundaries

Boundary	What it controls	Common trap
Azure management plane	Create/update/delete resources	Contributor on a resource does not always grant data-plane read/write.
Data plane	Use the service endpoint, indexes, models, or documents	Requires service-specific roles or keys.
Search index access	Query or modify indexes/documents	Separate query access from index administration.
Storage access	Read/write source documents	Indexers and apps need appropriate storage permissions.
Model deployment access	Invoke deployed models	Users/apps may need data-plane permission even if they can view the resource.
Application authorization	Which user can perform business action	Do not delegate this decision to the model.

Network Isolation Checklist

Requirement	Control
Keep traffic off public internet where supported	Private Endpoint / Private Link
Restrict public access	Disable or limit public network access and configure firewalls
Allow specific Azure services or networks	Service firewall rules and network integration
Resolve private endpoints correctly	Private DNS zone configuration
Secure app-to-service calls	Managed identity plus private endpoint where supported
Protect inbound app APIs	API Management, authentication, authorization, WAF where appropriate
Log security-relevant events	Azure Monitor, diagnostics, app logs, and audit trails

Responsible AI and Safety Controls

Concern	Practical control
Harmful content	Azure AI Content Safety, model content filters, blocked categories, review workflow
Hallucination	RAG grounding, citations, retrieval evaluation, refusal behavior when context is insufficient
Prompt injection	Treat retrieved/user text as untrusted, separate instructions from data, validate tool calls
Data leakage	Access filtering before retrieval, output redaction, least privilege, private networking
Bias or unfairness	Representative test sets, human review, metric tracking, documented limitations
Overreliance	Confidence indicators, citations, escalation paths, user education
Unsafe actions	Human-in-the-loop approval, allowlisted tools, transaction limits, audit logs
Privacy	Minimize collected data, redact where appropriate, control logs and retention

Deployment Architecture Patterns

Pattern	Use when	Azure services commonly involved
Synchronous AI API	User waits for response	App Service / Container Apps / AKS, Azure OpenAI, Azure AI Search, API Management
Event-driven document ingestion	Files arrive asynchronously	Blob Storage, Event Grid, Azure Functions, Azure AI Search, Document Intelligence
Long-running batch processing	Large document sets or audio transcription	Queue/Service Bus, Functions/Container Apps, durable workflow pattern, Storage
Chat application with RAG	Conversational answers over enterprise data	Web app, Azure OpenAI, Azure AI Search, storage, identity provider
Bot interface	Teams/web chat integration	Azure Bot Service, CLU/Language, Azure OpenAI, backend APIs
Custom ML endpoint	Model trained outside prebuilt AI services	Azure Machine Learning endpoint, app host, monitoring
Enterprise API facade	Standardized access to AI backend	API Management, managed identity, rate limiting, logging, backend services
Private enterprise deployment	Sensitive data and restricted access	Private Endpoints, VNet integration, managed identity, Key Vault, diagnostics

Hosting Service Selection

Need	Prefer	Notes
Simple web API or web app	Azure App Service	Good default for managed hosting.
Lightweight event handler	Azure Functions	Good for triggers, ingestion, and glue logic.
Containerized microservice without Kubernetes overhead	Azure Container Apps	Good for scalable container workloads and background workers.
Full Kubernetes control	AKS	Use when orchestration requirements justify complexity.
Workflow orchestration with connectors	Logic Apps	Good for integration-heavy business workflows.
Durable stateful orchestration	Durable Functions pattern	Good for fan-out/fan-in and long-running workflows.

Monitoring, Evaluation, and Optimization

What to Log and Monitor

Signal	Why it matters
Request count, latency, failure rate	Basic reliability and user experience
HTTP status codes	Diagnose auth, throttling, quota, and service errors
Token usage	Cost, latency, and prompt optimization
Model deployment used	Compare quality, regressions, and routing decisions
Prompt/template version	Reproduce failures and evaluate changes
Retrieval query and document IDs	Debug grounding and citation issues
Safety filter outcomes	Monitor blocked content and false positives/negatives
Tool calls and results	Audit actions and diagnose agent behavior
User feedback	Build evaluation datasets and prioritize fixes
Indexer status	Detect ingestion failures and stale search data

Evaluation Metrics by Scenario

Scenario	Evaluate
RAG answer generation	Groundedness, relevance, citation correctness, answer completeness, refusal when context is insufficient
Search retrieval	Precision, recall, top-k relevance, filter correctness, freshness
Classification	Accuracy, precision/recall, confusion matrix, threshold behavior
Extraction	Field-level accuracy, missing fields, table accuracy, format validity
Chat assistant	Task success, safety, latency, escalation rate, user satisfaction
Speech transcription	Word error patterns, domain vocabulary recognition, speaker/audio conditions
Document extraction	Model confidence, field accuracy, page/layout handling, exception routing

Optimization Levers

Problem	First levers to try
High latency	Smaller/faster model, shorter prompts, fewer retrieved chunks, caching, streaming responses, async processing
High cost	Token reduction, response length limits, model selection, cache repeated answers, batch offline tasks
Poor grounding	Improve chunking, add metadata, hybrid search, semantic ranking, better prompt constraints
Poor extraction	Use purpose-built service, add examples, validate schema, choose custom model when prebuilt fails
Frequent throttling	Backoff/retry, queue requests, smooth traffic, request capacity planning
Inconsistent output	Lower temperature, structured output, validation/retry, clearer instructions
Stale answers	More frequent indexing, event-driven ingestion, freshness filters

Troubleshooting Reference

Symptom	Likely cause	Check
401 Unauthorized	Missing/invalid credential	Endpoint, key/token, managed identity configuration
403 Forbidden	Authenticated but not authorized	RBAC role, data-plane permission, storage/search permissions
404 deployment not found	Wrong Azure OpenAI deployment name or endpoint	Resource endpoint, deployment name, region, API version
429 throttling	Too many requests or capacity pressure	Retry-after handling, exponential backoff, queueing, traffic smoothing
5xx service errors	Transient platform/backend issue	Retry with backoff, circuit breaker, monitor service health
Vector search returns no results	Wrong field, missing vectors, dimension mismatch, bad embeddings	Index schema, embedding model, vector field config, indexed documents
Answers hallucinate	Weak retrieval, prompt allows unsupported claims, no refusal rule	Retrieved chunks, citations, system instruction, evaluation set
Correct document not retrieved	Chunking/indexing/query mismatch	Chunk size, metadata filters, hybrid search, analyzers, synonyms
Citations are wrong	Source metadata not stored or chunk mapping incorrect	`sourceUri`, page/section fields, projection logic
Indexer fails	Data source permissions, unsupported file, skill error, mapping issue	Indexer execution history and skillset outputs
Private endpoint connection fails	DNS or network routing issue	Private DNS zone, VNet links, firewall, public access setting
Document fields missing	Prebuilt model mismatch or custom model undertrained	Document type, sample quality, confidence scores
CLU predicts wrong intent	Overlapping intents or weak utterance examples	Training data balance, labels, examples, thresholds
Speech transcription poor	Audio quality, noise, vocabulary, accent, wrong language	Audio preprocessing, custom speech, language config
Output JSON invalid	Model not constrained or schema too complex	Structured output, validation, repair/retry logic
Tool call unsafe or incorrect	Model over-selected tool or bad arguments	Tool allowlist, argument validation, user authorization

Common Exam Traps

Choosing Azure Machine Learning when a prebuilt Azure AI service already solves the task.
Choosing OCR when the requirement is structured form or table extraction; use Document Intelligence.
Treating a system prompt as a security control. It is guidance, not enforcement.
Letting the LLM decide whether a user is authorized to see retrieved content. Filter before retrieval or before prompt assembly.
Forgetting that Azure OpenAI apps call a deployment name, not just a model family.
Assuming embeddings generate answers. Embeddings support similarity search; a generative model writes the answer.
Ignoring metadata fields needed for filters, citations, freshness, and access control.
Using only vector search when exact IDs, codes, or names are important; consider hybrid search.
Assuming indexers run continuously. Understand scheduled, on-demand, and event-driven ingestion patterns.
Sending secrets, raw credentials, or excessive retrieved data into prompts.
Confusing management-plane RBAC with data-plane permissions.
Skipping retries and backoff for throttling and transient service failures.
Choosing the largest model automatically instead of balancing quality, latency, and cost.
Treating content filters as a full responsible AI program. They are one layer.

Scenario Answer Checklist

When you see an AI-200 scenario, identify:

Input type: text, image, document, audio, structured data, or mixed.
Task type: generate, retrieve, classify, extract, translate, transcribe, moderate, or train.
Data source: static, frequently updated, private, user-specific, or public.
Best service: prebuilt Azure AI service, Azure OpenAI, Azure AI Search, Azure ML, or a combination.
Security model: managed identity, RBAC, Key Vault, private endpoint, access trimming.
Runtime pattern: synchronous API, async queue, batch job, bot, web app, or containerized service.
Quality controls: evaluation set, citations, validation, thresholds, human review.
Operations: logging, monitoring, retries, throttling, cost/latency optimization.
Responsible AI controls: content safety, privacy, fairness, transparency, escalation.

Practical Next Step

Use this Quick Reference to build a one-page service-selection map, then practice scenario questions where you must justify the Azure AI service, security controls, retrieval pattern, and troubleshooting step. Focus especially on RAG design, managed identity/RBAC, Azure AI Search indexing, and choosing prebuilt AI services over custom models when appropriate.

Scenario Guide

Develop Containerized Solutions on Azure