Study Notes

Appearance

AI-102 Gap Analysis ​

Compared against the official AI-102 study guide (updated Dec 23, 2025).

Exam Structure Mismatch ​

The current notes are based on an older version of the AI-102 exam objectives. The exam was significantly restructured in 2025.

Official Domain (Current)WeightCurrent Notes
Plan and manage an Azure AI solution20-25%Domain 1 (listed as 15-20%)
Implement generative AI solutions15-20%Domain 5 (listed as 15-20%)
Implement an agentic solution5-10%COMPLETELY MISSING
Implement computer vision solutions10-15%Domain 3 (listed as 15-20%)
Implement natural language processing solutions15-20%Domain 4 (listed as 30-35%)
Implement knowledge mining and information extraction solutions15-20%Domain 6 (listed as 10-15%)

Domain 2: Content Moderation no longer exists as a separate domain. It has been folded into Domain 1 under "Implement AI solutions responsibly."

Template-only files with no actual AI-102 content: notes.md, objectives.md, quick-refresher.md, exam-tips.md.

Domain 1: Plan and Manage an Azure AI Solution (20-25%) ​

Covered Well ​

  • Resource types (single-service vs. multi-service)
  • Container deployment use cases (data residency, latency, billing metering)
  • Authentication methods (subscription keys vs. managed identities) with DefaultAzureCredential()
  • Network security (VNETs, Private Endpoints, Firewall Rules)
  • Diagnostic settings (Log Analytics, Storage, Event Hubs)
  • Alerts for common HTTP error codes (429, 5xx)

Missing or Too Shallow ​

  • [ ] Microsoft Foundry branding — Notes still use "Azure AI Services." Exam now uses "Microsoft Foundry Services," "Microsoft Foundry," "Azure AI Foundry." This affects how questions are worded.
  • [ ] Service selection guidance — Official objectives have an entire sub-section on "Select the appropriate Microsoft Foundry Services" for generative AI, vision, NLP, speech, information extraction, and knowledge mining. Need a decision tree or comparison table.
  • [ ] AI model selection and deployment options — Choosing the right AI models, standard vs. provisioned throughput, global vs. regional deployments.
  • [ ] CI/CD pipeline integration — "Integrate Microsoft Foundry Services into a CI/CD pipeline" is explicitly listed. Completely absent.
  • [ ] Cost management — "Manage costs for Microsoft Foundry Services" is tested. Only a brief mention in exam-guide, no detail.
  • [ ] Responsible AI governance framework — Exam expects "Design a responsible AI governance framework." Current notes only cover the 6 principles at surface level.
  • [ ] Prompt shields — Defending against prompt injection attacks. Not mentioned.
  • [ ] Content filters and blocklists for generative AI — Content filters applied to Azure OpenAI model outputs (input/output filters, configuring filter severity). Current Domain 2 covers Content Safety for user-generated content only.
  • [ ] Key Vault for key rotation — Mentioned as a one-liner in exam-guide but not explained.

Domain 2: Implement Generative AI Solutions (15-20%) ​

Covered Well ​

  • Chat Completions API roles (system, user, assistant)
  • Parameters (temperature, top_p, max_tokens, frequency/presence penalty)
  • Function calling concept
  • RAG concept and "On Your Data" feature with Azure AI Search and Blob Storage
  • Model list (GPT-4o, GPT-3.5 Turbo, Embeddings, DALL-E)

Missing or Too Shallow ​

  • [ ] Azure AI Foundry (hubs, projects) — The entire Foundry portal concept is absent. Need to cover:
    • What a Foundry hub is (shared infrastructure)
    • What a Foundry project is (workspace within a hub)
    • How to create and manage these resources
  • [ ] Prompt Flow — Visual tool in Azure AI Foundry for building LLM workflows. Need to cover:
    • Flow types (standard, chat, evaluation)
    • Nodes, connections, variants
    • Deploying flows as endpoints
  • [ ] Model evaluation — Not covered. Need:
    • Built-in evaluation metrics (groundedness, relevance, coherence, fluency)
    • Custom evaluation flows
    • Batch evaluation runs
  • [ ] Microsoft Foundry SDK — "Integrate your project into an application with Microsoft Foundry SDK." Not mentioned.
  • [ ] Prompt templates — Parameterized prompts for reuse. Not covered.
  • [ ] DALL-E image generation — Listed in cheatsheet but no implementation details (API call structure, image sizes, quality parameters).
  • [ ] Large multimodal models (LMMs) — GPT-4o/GPT-4 Vision processing images alongside text. Not covered.
  • [ ] Prompt engineering techniques — Need:
    • Few-shot prompting
    • Chain-of-thought
    • System message design patterns
    • Grounding techniques
  • [ ] Fine-tuning — Completely absent. Need:
    • When to fine-tune vs. use RAG
    • Training data preparation (JSONL format)
    • Fine-tuning workflow in Azure OpenAI
  • [ ] Model reflection — Technique where the model reviews and improves its own outputs.
  • [ ] Tracing and feedback collection — Azure AI Foundry tracing for debugging LLM calls and user feedback mechanisms.
  • [ ] Orchestration of multiple generative AI models — Chaining or routing between different models.
  • [ ] Content filters for Azure OpenAI — Default filters, custom filter configurations, annotation responses. Only a one-liner in exam-guide.
  • [ ] Container deployment for generative models on edge — Listed in objectives, not in notes.
  • [ ] Scalability and model updates — "Optimize and manage resources for deployment, including scalability and foundational model updates."

Domain 3: Implement an Agentic Solution (5-10%) — ENTIRELY MISSING ​

This is a completely new domain. Nothing exists in the current notes. Topics to cover:

  • [ ] Foundry Agent Service — What it is, how it orchestrates tool calls, manages conversations, enforces content safety.
  • [ ] Microsoft Agent Framework — Open-source framework for building agentic applications.
  • [ ] Agent concepts — Tools/functions, code interpreter, file search, Azure AI Search integration.
  • [ ] Multi-agent orchestration — Patterns for multiple specialized agents collaborating.
  • [ ] Agent testing and deployment — How to test, optimize, and deploy agents.
  • [ ] Autonomous capabilities — How agents can execute multi-step tasks independently.

References ​

Domain 4: Implement Computer Vision Solutions (10-15%) ​

Covered Well ​

  • Image Analysis 4.0 features (captioning, dense captioning, tagging, smart cropping, people detection)
  • Custom Vision classification types (multiclass vs. multilabel) and object detection
  • Training loop (upload, tag, train, evaluate, publish, test)
  • Face API capabilities (detection, verification 1:1, identification 1:N)
  • Limited Access policy and retired features (emotion, gender/age)
  • OCR Read API async pattern (202 -> operation-location -> GET)

Missing or Too Shallow ​

  • [ ] Azure Vision in Foundry Tools — Branding change. "Azure Vision in Foundry Tools" replaces "Azure AI Vision."
  • [ ] Video Indexer — Extracting insights from video/live streams:
    • Faces, topics, sentiments, brands, scenes, keyframes
    • OCR in video
  • [ ] Spatial Analysis — Detecting presence and movement of people in video feeds:
    • People counting
    • Distance monitoring
    • Zone dwell time
  • [ ] Handwritten text extraction — Exam distinguishes "Convert handwritten text using Azure Vision in Foundry Tools." Notes mention OCR/Read but not handwriting specifically.
  • [ ] Code-first custom vision — "Build a custom vision model code first" is a new objective. Notes focus on portal-based workflow only.
  • [ ] Custom vision model metrics — mAP (mean Average Precision) for object detection, evaluation dashboard interpretation.
  • [ ] Face API depth — Missing:
    • PersonGroup / LargePersonGroup management workflow (create group -> add persons -> add faces -> train -> identify)
    • FaceList / LargeFaceList for find-similar operations
    • Face detection attributes still available (blur, exposure, noise, accessories, head pose, occlusion)

Domain 5: Implement Natural Language Processing Solutions (15-20%) ​

Covered Well ​

  • Language Service features (sentiment, key phrases, entity linking, PII, language detection)
  • Custom NER and Custom Text Classification mentioned
  • CLU workflow (schema, label, train, test, publish)
  • Speech-to-Text (real-time, batch, custom speech)
  • Text-to-Speech (neural voices, SSML, custom neural voice)
  • Translator service (text, document, custom translator with TMX)
  • Transliterate mentioned in cheatsheet
  • Precision/Recall/F1 metrics in cheatsheet

Missing or Too Shallow ​

  • [ ] Custom Question Answering (formerly QnA Maker) — Major gap. Exam has 7 sub-objectives:
    • Create a custom question answering project
    • Add question-and-answer pairs and import sources
    • Train, test, and publish a knowledge base
    • Create multi-turn conversations
    • Add alternate phrasing and chit-chat
    • Export a knowledge base
    • Create a multi-language question answering solution
  • [ ] Intent and keyword recognition with Speech — Using Speech SDK to recognize specific keywords as wake words and detect user intent from spoken input.
  • [ ] Speech translation — Speech Translation API (real-time speech-to-text translation, speech-to-speech translation). Notes cover text translation only.
  • [ ] Generative AI speaking capabilities — "Integrate generative AI speaking capabilities in an application." Combining Azure OpenAI with Speech services.
  • [ ] CLU model management — "Optimize, backup, and recover language understanding model." Exporting/importing CLU models, versioning, deployment slots.
  • [ ] Custom translation depth — Training, improving, and publishing a custom translation model in detail.
  • [ ] Entity linking vs. NER — Need clearer distinction between Entity Linking (Wikipedia disambiguation) and standard Named Entity Recognition.

Domain 6: Implement Knowledge Mining and Information Extraction (15-20%) ​

Covered Well ​

  • AI Search architecture (data source, indexer, skillset, index)
  • Semantic Search and Vector Search concepts
  • Document Intelligence models (Read, Invoice, Receipt, ID, W-2)
  • Custom models (Template vs. Neural) and Composed Models
  • Knowledge Store mentioned in cheatsheet
  • Enrichment pipeline diagram

Missing or Too Shallow ​

  • [ ] Azure Content Understanding in Foundry Tools — Completely new service. Need:
    • Create an OCR pipeline to extract text from images and documents
    • Summarize, classify, and detect attributes of documents
    • Extract entities, tables, and images from documents
    • Process and ingest documents, images, videos, and audio
  • [ ] Custom skills in Azure AI Search — Need:
    • Custom Web API skills (calling Azure Functions)
    • Custom skill interface (input/output schema)
    • Built-in vs. custom skills distinction
  • [ ] Query syntax depth — Need:
    • Lucene query syntax (simple vs. full)
    • $filter, $orderby, $select, $top, $skip OData parameters
    • Wildcard and regex queries
    • Faceted navigation
  • [ ] Knowledge Store projections detail — Need:
    • Table projections, object projections, file projections
    • Shaper skill for structuring data before projection
    • Use cases (analytics in Power BI, secondary processing)
  • [ ] Document Intelligence training workflow — Need:
    • Labeling tool (Document Intelligence Studio)
    • Minimum training document requirements
    • Model evaluation metrics (accuracy, confidence scores)
  • [ ] Semantic and vector search depth — Need:
    • Vector index configuration (HNSW algorithm, dimensions, metric)
    • Hybrid search (combining keyword + vector)
    • Integrated vectorization (built-in embedding skill)

Priority Summary ​

P0 — Must Add (Completely Missing, High Exam Impact) ​

GapExam Weight
Agentic Solutions domain5-10%
Custom Question AnsweringPart of NLP (15-20%)
Azure AI Foundry (hubs, projects, Prompt Flow)Part of GenAI (15-20%)
Azure Content UnderstandingPart of Knowledge Mining (15-20%)
Video Analysis (Video Indexer + Spatial Analysis)Part of Vision (10-15%)

P1 — Must Deepen (Covered but Insufficient) ​

GapDomain
Prompt engineering techniquesGenAI
Fine-tuning generative modelsGenAI
AI Search query syntaxKnowledge Mining
Custom skills in AI SearchKnowledge Mining
Content filters for Azure OpenAIGenAI
Knowledge Store projectionsKnowledge Mining
Speech translationNLP
CI/CD for AI servicesPlan & Manage
Responsible AI governance frameworkPlan & Manage
Face API (PersonGroup workflow)Vision
Document Intelligence trainingKnowledge Mining

P2 — Must Update (Branding/Structure) ​

ItemChange Needed
Domain numbering and weightsRestructure to match 6 official domains
Service names"Azure AI Services" -> "Microsoft Foundry Services"
Remove standalone Domain 2Merge Content Moderation into Domain 1
Template filesFill notes.md, objectives.md, quick-refresher.md, exam-tips.md