AI-102 Consolidated Study Notes

🛠️ Microsoft AI Foundry Architecture

• Foundry Hub: Common resource for security, compute, and connections.
• Foundry Project: Collaborative workspace for building AI apps.
• Model Catalog: One-stop shop for Azure OpenAI, Meta, Mistral, and other models.

🤖 Generative AI Implementation

• Chat Completions API:

  from azure.ai.inference import ChatCompletionsClient
  
  client = ChatCompletionsClient(endpoint, credential)
  response = client.complete(
      messages=[
          {"role": "system", "content": "You are a help assistant."},
          {"role": "user", "content": "How do I use Prompt Flow?"}
      ],
      temperature=0.7
  )

• RAG Pattern: Retrieve (Azure AI Search) -> Augment (Insert as context) -> Generate (LLM).
• Prompt Flow: Visual tool with LLM, Python, and Prompt nodes.

🕵️ Agentic Solutions

• Agent Service: Uses a reasoning loop (CoT) to decide which tools to call.
• Tool Definition: Describing functions in JSON schema so the model knows how to "invoke" them.
• Sandboxed Execution: Using Code Interpreter for safe Python execution.

👁️ Vision & NLP

• OCR Pattern: POST to /analyze -> Get Operation-Location -> Poll until succeeded.
• Face Identification: 1:N search against a PersonGroup.
• CLU: Map user utterances to Intents. Use prebuilt entities for dates/numbers.
• QA: Import data sources (URLs, PDFs) to create a conversational KB.

🔍 Search & Mining

• Skillsets: Chain of AI skills (OCR, KeyPhrases, Custom Web API).
• Hybrid Search: Keyword + Vector (HNSW) + Semantic Re-ranker.
• Document Intelligence: Layout, Prebuilt, and Custom (Template vs Neural) models.