AI-103 Course Notes

Planning an AI-103 solution starts with choosing the right retrieval and model services. Azure AI Search is the stronger fit when the app needs a search index with vector fields, hybrid keyword plus vector retrieval, filters, semantic ranking, and scalable relevance tuning. Cosmos DB is a better fit when the app already needs an operational NoSQL store for transactional data and lightweight vector lookup over that same data; it is not a drop-in replacement for a full search engine with analyzers, semantic ranker, and indexer pipelines.

• Choosing Cosmos DB when the scenario really needs Azure AI Search features such as hybrid retrieval, semantic ranking, analyzers, or indexer pipelines.

• Vector plus keyword retrieval should make Azure AI Search hybrid search stand out.

Extracting structured data from documents is a core capability of Azure AI Document Intelligence. The process begins with selecting the appropriate model type - either a prebuilt model for standard document layouts or a custom neural model for unstructured documents. The Layout model is particularly useful for extracting data from forms and structured documents, while Custom Neural Models provide flexibility for variable layouts. Asynchronous analysis enables non-blocking document processing, allowing applications to retrieve extracted data once the Document Intelligence service completes its analysis.

• Choosing a custom model when a prebuilt Document Intelligence model already matches the document type.

• Use prebuilt Invoice for standard invoice fields and prebuilt ID for identity documents.

Azure AI Vision is the service choice when the app needs computer-vision features exposed through Azure AI services, such as image captions, dense captions, tags, objects, optical character recognition, and other Image Analysis API visual features. The exam often contrasts this with GPT-4o Vision: use Azure AI Vision for structured visual feature extraction and service-specific image analysis, and use GPT-4o Vision when the task needs multimodal reasoning over image content with natural-language context.

• Choosing GPT-4o Vision when the requirement is fixed Image Analysis API features such as tags, captions, OCR, or objects.

• Use Azure AI Vision for structured image features from the Image Analysis API.

Prompt Flow and Semantic Kernel serve different implementation styles. Prompt Flow is best when a team wants a visual or flow-oriented way to build, evaluate, and debug prompt pipelines in Foundry. Semantic Kernel is a code-first SDK for connecting prompts, functions, plugins, memory, and agents inside an application. The exam distinction is design and evaluation flow versus programmatic orchestration.

• Using Prompt Flow and Semantic Kernel interchangeably instead of separating flow evaluation from code-first orchestration.

• Prompt Flow fits visual pipeline design, evaluation, and debugging.

Administrators configure Azure AI Content Filtering to mitigate risk by defining category-specific filtering rules. These rules, combined with severity thresholds, determine the action taken when content matches a category. Administrators monitor and adjust these settings through the Azure portal, utilizing Log Analytics to track token usage and evaluate model performance. Fine-tuning involves adjusting thresholds based on observed results and leveraging the Evaluator Prompt Criteria to refine groundedness evaluations.

• Changing the whole content filter policy when only a severity threshold or blocklist needs tuning.

• Use content filtering and blocklists for policy enforcement around generated content.

Establishing consistent Generative AI outputs begins with defining a clear system prompt. This prompt, delivered as a high-priority instruction set, guides the Azure OpenAI chat model's responses, shaping its role, tone, and the type of information it generates. Utilizing system prompts allows administrators to control the model's behavior and ensure outputs align with brand guidelines. The System Prompt Persona control within the Azure portal enables the creation of distinct personas for the model, further refining its responses based on specific requirements.

• Trying to enforce persona with repeated user prompts instead of a system message or reusable prompt template.

• Use system messages for stable behavior, persona, constraints, and response style.

Secure AI solution design starts with identity. Managed identity lets Azure OpenAI, apps, functions, or other Azure resources authenticate to services without storing keys in code. Key Vault should hold secrets and keys, and Azure RBAC should grant the managed identity only the roles it needs. A 401 error in this area usually means authentication succeeded nowhere or authorization is missing for the exact data source, Key Vault secret, Search index, or API being called.

• Fixing a 401 by changing network settings when the real problem is missing managed identity or data-plane RBAC.

• Separate identity, network, and permission troubleshooting before choosing an answer.

A RAG architecture retrieves trusted content and supplies it to a model as grounding context. Azure AI Search commonly handles chunked content, vector fields, keyword fields, filters, semantic ranking, and indexer-driven ingestion. The model answers from the retrieved context, but grounding quality depends on chunking, embedding model choice, vector dimensions, query design, ranking, and prompt instructions such as strictness and topNDocuments.

• Creating a vector field with dimensions that do not match the embedding model output.

• RAG quality depends on chunking, embeddings, vector schema, retrieval, ranking, and prompt grounding.