AI-102 study guide
Key facts by service. AI-102 goes a level deeper than AI-900: SDK and REST specifics, resource configuration, security and the order of operations in a pipeline.
Key facts by topic
Azure AI Search
| Concept | Key Facts |
|---|---|
| Index Properties | filterable=WHERE, sortable=ORDER BY, facetable=counts, retrievable=include in results |
| Indexer Setup Order | Data Source → Index → Indexer → Schedule |
| Indexer Properties | skillsetName + outputFieldMappings for enrichments |
| REST Operations | Run=execute now, Reset=full re-index next run, Update=change definition |
| Security | Separate data sources for folders, Entra ID security filters for trimming |
| Query API Keys | Use management.azure.com endpoint + Microsoft.Search provider + createQueryKey |
| Hybrid Search | BM25 + vector scoring for best keyword + semantic relevance |
| Partitions vs Replicas | Partitions=query speed, Replicas=concurrent load (2 replicas=99.9% SLA read-only) |
Knowledge Store
| Projection Type | Storage | Use Case |
|---|---|---|
| Tables | Azure Table Storage | Transactional/tabular data, structured rows |
| Objects | Azure Blob Storage | JSON documents from enrichments |
| Files | Azure Blob Storage | Binary files, images (JPEG, PNG) |
| Property | Purpose |
|---|---|
| source | Path to enrichment (e.g., /document/tableprojection/keyPhrases/*) |
| tableName | Azure Table storage table name |
| storageContainer | For blob/object projections |
Azure OpenAI / Generative AI
| Concept | Key Facts |
|---|---|
| Chat Completion API | client.chat.completions.create() - GPT-4, GPT-4o, o3 models |
| Completion API | client.completions.create() - GPT-35-Turbo-Instruct only |
| Model Parameter | = deployment name in Azure OpenAI (not base model name) |
| Fine-tuning Format | JSONL (JSON Lines) only - NOT XML, JSON, or CSV |
| Fine-tuning Steps | Convert to JSONL → Upload to Blob Storage → POST request |
| Embedding Model | text-embedding-ada-002 (NOT GPT-4) |
| Multimodal Images | Base64 data URI or URL with SAS token in image_url field |
| Usage Object | prompt_tokens + completion_tokens + total_tokens |
Azure AI Foundry & Agents
| Concept | Key Facts |
|---|---|
| Agent Components | Language Model + Instructions + Tools |
| Blocklists | ai.azure.com → Safety + security → Blocklists tab → Add blocklist |
| Agent Tools | Bing Search=public web, Logic Apps=external workflows, Code Interpreter=Python |
| File Upload | Correct connection string + Storage Blob Data Owner role on agents-blobstore |
| Conversation Access | Full access to structured threads (user-agent and agent-agent) |
| Data Storage | Does NOT auto-manage all data - requires external connections |
Custom Vision
| Concept | Key Facts |
|---|---|
| Use Case | Train your own model for custom objects (logos, animals, products) |
| Portal | customvision.ai (NOT Azure portal) |
| Object Detection | Returns location (bounding boxes) + classification |
| Classification | Labels only, no location info |
| Edge Deployment | Use Compact domains (General compact, etc.) |
| Training Method | TrainProject() - returns iteration object |
| Polling Status | iteration.Status == "Training" while in progress |
| TensorFlow Export | model.pb (ProtoBuf) + labels.txt |
| Metrics API | GetIterationPerformance (not GetImagePerformances) |
| Iteration | Doesn't guarantee improvement - need more/better training data |
Azure AI Vision
| Concept | Key Facts |
|---|---|
| Endpoint Format | {resourceName}.cognitiveservices.azure.com |
| visualFeatures | Tags=word list, Categories=86-class taxonomy, Adult=nudity/violence, Brands=logos |
| details | Celebrities, Landmarks |
| OCR Polling | GetReadResultAsync, check Running AND NotStarted status |
| vs Custom Vision | Azure AI Vision = pre-trained; Custom Vision = train your own |
Document Intelligence
| Concept | Key Facts |
|---|---|
| Use Case | Extract key-value pairs, text, tables from forms/documents |
| Custom Template | Rule-based, fixed layouts, consistent format |
| Custom Neural | Deep learning, mixed layouts, format variations |
| Composed Models | Group existing models → one model ID → auto-classification |
| Export Format | JSON only (not XML) |
| Access Methods | Document Intelligence Studio + SDKs (C#, Python, Java, JavaScript) |
| Min Training Docs | 5 documents for custom models |
| vs NER | Document Intelligence = structured extraction; NER = entity recognition in text |
Azure AI Language
| Feature | Key Facts |
|---|---|
| CLU Utterance | User input text (the "text" field) |
| CLU Intent | What user wants to do (action category) |
| CLU Offset | Character position where entity starts |
| CLU Training Split | Manual split = control specific utterances; Auto = percentage-based only |
| CLU Failover | Create resource FIRST → Export → Import → Train/Deploy |
| Language Detection | countryHint defaults to "US", stateless service, confidence 0-1 |
| Sentiment "Mixed" | Requires at least one positive AND one negative sentence |
| QA Refresh URL | Updates from source, retains manual edits |
| QA Alternative Questions | Different phrasings to improve search |
| QA Follow-up Prompts | Guide users through conversation path |
| QA Metadata | Categorize/tag Q&A pairs |
| KB Export | Edit knowledge base (not Deploy) → Export as TSV |
Speech Services
| Concept | Key Facts |
|---|---|
| recognize_once() | Max 15 seconds, single operation |
| start_continuous_recognition() | Unlimited duration, ongoing |
| SpeechRecognitionLanguage | Source language (what you're listening for) |
| AddTargetLanguage | Translation output languages |
| PhraseListGrammar.addPhrase | Improve recognition of specific terms |
| SSML | Adjust pitch, rate, pronunciation, volume |
| AudioDataStream | Streaming for lower latency |
| Tuning file | Audio Content Creation tool for fine-tuning |
| Batch Transcription | Speech to Text REST API or Azure Speech CLI (not SDKs) |
| Whisper Limit | 25MB file size maximum |
| Custom Keyword | Basic (minutes, demos) vs Advanced (48 hours, production) |
| Visemes | Lip-sync animation (position of lips, jaw, tongue) |
Content Safety
| Concept | Key Facts |
|---|---|
| Content Filters | Block harmful content (hate, sexual, violence) |
| Prompt Shields | Detect adversarial prompts (jailbreaks, prompt injections) |
| Prompt Shields Scope | Scans user prompts + documents (RAG sources), NOT model outputs |
| AnalyzeTextOptions | Class for creating Content Safety request (not AnalyzeTextOp) |
| Package | azure.ai.contentsafety (sync), azure.ai.contentsafety.aio (async) |
| Export | Filter config exportable to Python, Java, C# |
Semantic Kernel
| Concept | Key Facts |
|---|---|
| Concurrent Pattern | Parallel independent tasks |
| Sequential Pattern | Ordered pipeline/workflow |
| Handoff Pattern | Dynamic escalation/transfer of control |
| OpenAPI Plugin | Auto-generates functions from Swagger/OpenAPI spec |
Containers & Infrastructure
| Concept | Key Facts |
|---|---|
| Docker Port Mapping | -p host:container (e.g., -p 8080:5000) |
| Container Params | Eula=accept, Billing=endpoint URL, ApiKey=API key |
| Key Rotation | Switch to Key2 → Reset Key1 → Switch back to Key1 |
Observability
| Concept | Key Facts |
|---|---|
| Tracing Implementation | "opentelemetry" (not logging.DEBUG) |
| Console Export | ConsoleSpanExporter() for STDOUT |
| Azure Monitor | AzureMonitorTraceExporter (needs connection string) |
| OTLP Export | OTLPSpanExporter (needs endpoint) |
More key facts
Azure AI Search
| Concept | Key Facts |
|---|---|
| Indexing Pipeline | Document Cracking → Field Mappings → Skillset Execution → Output Field Mappings → Push into Index |
| Partitions | Split data across compute - improves individual query speed |
| Replicas | Load balancing - handles more concurrent users |
| Indexer Setup | Create data source → Define index → Create indexer → Schedule indexer |
| OCR Skillset | Only needed for images, not database content |
Azure OpenAI
| Concept | Key Facts |
|---|---|
| Use Case | Creative content generation (descriptions, stories, code, summaries) |
| Fine-tuning Format | JSONL (JSON Lines) - NOT XML, JSON, or CSV |
| Fine-tuning Steps | Convert to JSONL → Upload to Blob Storage → POST to fine_tuning/jobs |
| REST API | Synchronous responses - NO webhooks needed |
| API Integration | Get endpoint/key → Build payload → POST request → Parse JSON response |
DALL-E (Azure OpenAI)
| Feature | Details |
|---|---|
| Output Format | PNG only (not JPEG) |
| Image Sizes | 1024x1024, 1024x1792, 1792x1024 |
| View Code | Yes - prefilled Python code available in playground |
Azure AI Document Intelligence
| Concept | Key Facts |
|---|---|
| Use Case | Extract key-value pairs, text, tables from forms/documents |
| Pre-built Models | Invoices, receipts, tax forms, IDs |
| Custom Models | Train for unique document types (min 5 training docs) |
| Composed Models | Group multiple models → single model ID → auto-classification |
| NOT for | General text analysis (use Azure AI Language instead) |
Azure AI Language
| Concept | Key Facts |
|---|---|
| Use Case | Analyze and understand text |
| Features | NER, sentiment analysis, key phrases, summarization, language detection |
| NOT for | Form extraction (use Document Intelligence) |
| NOT for | Creative content generation (use Azure OpenAI) |
| CLU Flat Architecture | Max ~50 intents, single feature |
| CLU Orchestration | Multiple child apps + orchestration for 50+ intents |
Azure AI Vision
| Concept | Key Facts |
|---|---|
| Endpoint Format | https://{region}.api.cognitive.microsoft.com/vision/v2.0/analyze |
| Region Name | Use exact name: westus, eastus (not westu, not resource name) |
| visualFeatures | Description, Tags, Categories, Objects, Faces, etc. |
| details | Landmarks, Celebrities |
| OCR Methods | GetReadResultAsync for read results, check Running AND NotStarted status |
Azure AI Speech
| Method | Max Duration | Use Case |
|---|---|---|
| recognize_once() | 15 seconds | Short utterances |
| recognize_once_async() | 15 seconds | Short utterances (async) |
| start_continuous_recognition() | Unlimited | Long audio (10s - minutes) |
| addPhrase() | N/A | Improve recognition of custom terms |
Azure AI Foundry
| Concept | Key Facts |
|---|---|
| Hub Security | Settings inherited by projects |
| Project Region | MUST be same region as parent hub |
| Cross-region Connections | Projects CAN connect to Azure OpenAI in different regions |
Azure AI Containers (Disconnected)
| Source | Supported? |
|---|---|
| Azure AI Foundry model catalog | No |
Knowledge Mining
| Phase | Description |
|---|---|
| 1. Ingest | Import and aggregate structured/unstructured data |
| 2. Enrich | Apply AI skills (OCR, NER, key phrases, translation) |
| 3. Explore | Make searchable, visualize, discover patterns |
Cost Management
| Alert Type | Triggers When |
|---|---|
| Actual Cost Alert | Actual spending reaches % threshold |
| Forecasted Alert | ML predicts spending will exceed % threshold |
| Budget Reset | Monthly budgets reset at end of each month |