AI-102 cheat sheet
Decision rules, key numbers and memory aids for the Azure AI services AI-102 covers.
Decision rules
Content Generation:
- Generate creative text/descriptions? → Azure OpenAI
- Analyze existing text (sentiment, entities)? → Azure AI Language
- Extract data from forms? → Azure AI Document Intelligence
Custom Vision vs Azure AI Vision:
- Need to train on YOUR specific images? → Custom Vision
- Using pre-built general algorithms? → Azure AI Vision
- Need bounding boxes/locations? → Object Detection project
- Just need labels? → Classification project
Document Intelligence Model Type:
- Fixed layouts, consistent format? → Custom Template
- Mixed layouts, format variations? → Custom Neural
- Common forms (invoices, receipts)? → Prebuilt models
- Multiple form types, one model ID? → Composed model
Azure AI Search Performance:
- Individual queries slow? → Add partitions
- Too many concurrent users? → Add replicas
- Improve performance? → Reduce fields, use simpler types
Speech Recognition:
- Audio < 15 seconds? → recognize_once()
- Audio > 15 seconds? → start_continuous_recognition()
- Improve custom term recognition? → addPhrase()
- Large audio files batch processing? → REST API or CLI (not Whisper for >25MB)
CLU Training Data:
- Control which specific utterances go where? → Manual split
- Just need percentage-based division? → Automatic split
Knowledge Store Projections:
- Tabular/transactional data? → Tables
- JSON documents? → Objects
- Images/binary files? → Files
Search Types:
- Exact keyword matching? → Keyword search (BM25)
- Semantic understanding? → Vector search (uses embedding model)
- Best of both? → Hybrid search
- Re-rank for relevance? → Semantic ranking (post-processing)
API Types:
- Chat conversation format? → Chat Completion API
- Instruction/command following? → Completion API (GPT-35-Turbo-Instruct)
- Generate embeddings? → Embedding API (text-embedding-ada-002)
Key numbers
| Number | Meaning |
|---|---|
| 25 MB | Whisper model file size limit |
| 15 seconds | Max duration for recognize_once() |
| 48 hours | Advanced custom keyword model training time |
| 0-1 | Confidence score range (not 0-100) |
| 5 | Minimum training documents for custom Document Intelligence model |
| 2 replicas | Required for 99.9% SLA on read-only workloads |
| 3+ replicas | Required for high availability on read-write workloads |
| 86 | Number of classes in Azure AI Vision Categories taxonomy |
Memory Aids
Knowledge Store Projections (TFO)
- Tables = Tabular data (Table Storage)
- Files = Files/images (Blob Storage)
- Objects = Output JSON (Blob Storage)
Index Properties (FSFR)
- Filterable = Filter results
- Sortable = Sort order
- Facetable = Facet counts
- Retrievable = Return in results
Search Indexer Setup (DISC)
- Data source
- Index definition
- Set up indexer (create)
- Configure schedule
Fine-tuning Steps (CUP)
- Convert to JSONL
- Upload to Blob Storage
- POST to submit job
Key Rotation (SRS)
- Switch to backup key
- Reset compromised key
- Switch back to primary
RAG Evaluation Order (SPEARM)
- Select target
- Prepare test data
- Evaluators (identify)
- Analyze results (after running)
- Run evaluation
- Make revisions
CLU Failover (CEIT)
- Create resource in target region
- Export project
- Import to new resource
- Train and deploy
Knowledge Mining (IEE)
- Ingest and aggregate
- Enrich with AI
- Explore and discover
Speech Recognition Duration
- once = 15 seconds max (short)
- continuous = unlimited (long)
Custom Vision Projects
- Detection = Dimensions (bounding boxes)
- Classification = Categories only