AI-102 Cheatsheet (2025)
Core Infrastructure (Domain 1)
| Concept | Key Detail |
|---|---|
| AI Foundry Hub | Shared compute, connections, security — for multiple teams |
| AI Foundry Project | Your workspace inside a Hub — build, test, deploy |
| Managed Identity | DefaultAzureCredential() — keyless auth for production |
| PTU (Provisioned) | Reserved capacity, predictable latency, fixed cost |
| Standard | Pay-per-token, variable latency, shared infrastructure |
| Docker Container | Edge/offline/data residency deployments |
| Content Safety | Standalone service: Hate / Self-Harm / Violence / Sexual, 4 severity levels |
| Prompt Shields | Blocks prompt injection attacks |
| Content Filters | Applied to Azure OpenAI model inputs AND outputs |
Generative AI (Domain 2)
| Concept | Key Detail |
|---|---|
| Chat roles | system (instructions) → user (prompt) → assistant (history) |
| temperature | 0 = deterministic; higher = more creative/random |
| RAG | Inject source docs at runtime; fast, updatable, no retraining |
| Fine-tuning | Bake tone/format into weights; requires JSONL training data (100–500+ examples) |
| Prompt Flow | Visual LLM workflow builder — Standard / Chat / Evaluation flow types |
| Variants | A/B test different prompt versions within one Prompt Flow |
| Groundedness | RAG quality metric — does the response stay in the source context? |
RAG vs Fine-tuning
| Factor | RAG | Fine-tuning |
|---|---|---|
| Data changes often | Use | Avoid |
| Need specific tone/format | Partial | Use |
| Reduce hallucinations | Use | Partial |
| Requires retraining | No | Yes |
Agentic Solutions (Domain 3)
| Concept | Key Detail |
|---|---|
| AI Agent Service | Managed service for autonomous multi-step reasoning |
| Thread | Conversation session — stores message history |
| Run | One execution of the reasoning loop on a Thread |
| Code Interpreter | Agent writes + executes Python in a secure sandbox |
| File Search | Agent retrieves from uploaded documents (built-in vector search) |
| Multi-agent: Hierarchical | Manager delegates to Worker agents |
| Multi-agent: Sequential | Pipeline — each agent passes output to the next |
| Multi-agent: Group Chat | Agents discuss and critique together |
Computer Vision (Domain 4)
| Concept | Key Detail |
|---|---|
| Image Analysis 4.0 | Captioning, Dense Captioning, Tagging, Smart Crop, People Detection |
| Read API (OCR) | Async: POST → 202 + Operation-Location → GET until succeeded |
| Custom Vision — Classification | Multiclass (1 tag) or Multilabel (multiple tags) per image |
| Custom Vision — Object Detection | Returns tag + bounding box coordinates |
| Video Indexer | Semantic insights from video: faces, topics, brands, transcript, sentiment |
| Spatial Analysis | Real-time movement in live feed — runs on edge Docker container |
| Face Verification | 1:1 — are these two faces the same person? |
| Face Identification | 1:N — who is this from a PersonGroup? (requires training) |
| FaceList | 1:N Find Similar — no person concept, no training needed |
NLP and Speech (Domain 5)
| Concept | Key Detail |
|---|---|
| Sentiment Analysis | Document + sentence level: positive/negative/neutral |
| NER | Extracts entity types (Person, Location, Date, etc.) |
| Entity Linking | Disambiguates entities via Wikipedia |
| PII Detection | Detects and can redact personal data |
| CLU | Utterance → Intent + Entities; replaces LUIS |
| Custom QA | Matches questions to stored Q&A pairs; supports multi-turn, chit-chat |
| Speech STT | Real-time and batch transcription |
| Intent Recognition | CLU intent detection from spoken audio via Speech SDK |
| Keyword Recognition | Local on-device wake word detection — no cloud needed |
| Document Translation | Async — translates Word/PDF preserving layout |
| Custom Translator | Domain-specific translation using TMX/XLIFF parallel corpora |
CLU vs Custom QA
| CLU | Custom QA | |
|---|---|---|
| Input | Utterance | Question |
| Output | Intent + Entities | Answer text |
| Use case | Take an action | Return information |
| Training | Label utterances | Import Q&A sources |
Knowledge Mining (Domain 6)
| Concept | Key Detail |
|---|---|
| AI Search Pipeline | Data Source → Indexer → Skillset → Index |
| Custom Skill | External Azure Function — must follow values[] input/output schema |
| Knowledge Store | Table (Power BI) / Object (blob JSON) / File projections |
| Shaper Skill | Reshapes enriched data for Knowledge Store projections |
| Hybrid Search | Keyword (BM25) + Vector (embeddings) + RRF score fusion |
| Semantic Ranking | LLM re-ranker applied after retrieval — surfaces best answer |
| HNSW | Algorithm used for vector similarity search |
| Content Understanding | New multimodal pipeline (docs + images + video + audio) |
| Doc Intelligence: Template | Fixed-layout forms — 5+ training docs |
| Doc Intelligence: Neural | Variable-layout docs (contracts) — 100+ training docs |
| Doc Intelligence: Composed | Routes to the best matching custom model |
AI Search Query Parameters
| Parameter | Purpose |
|---|---|
$filter | Boolean OData filter (e.g., Category eq 'Finance') |
$select | Return specific fields only |
$top / $skip | Pagination |
$orderby | Sort results |
Async Patterns (High Frequency Exam Topic)
All three follow the same 202 → Operation-Location → GET pattern:
| Operation | Service |
|---|---|
| OCR / Read API | Azure AI Vision |
| Document Translation | Azure AI Translator |
| Batch operations | Azure OpenAI, Document Intelligence |