初學者專用生成式AI
一門全面的課程，探討生成式AI、大型語言模型、提示工程的基礎知識，以及使用Azure OpenAI和Power Platform等工具開發AI應用程式的技巧。
課程總覽
📚 內容摘要
一個全面的課程，探討生成式AI、大型語言模型、提示工程，以及使用Azure OpenAI和Power Platform等工具開發人工智慧應用程式的基礎知識。
掌握生成式AI的基本原理，從零開始打造智慧型應用程式。
致謝： Microsoft、Azure 和 OpenAI。
🎯 學習目標
- 解釋大型語言模型（LLM）的機械內部運作原理，包括分詞、注意力機制及非確定性輸出。
- 比較不同類型的大型語言模型（基礎模型、開源與專有模型，以及編碼器/解碼器架構），以針對業務情境選擇最合適的工具。
- 評估提升模型效果的策略，特別是判斷應選擇提示工程、檢索增強生成（RAG）或微調。
- 定義提示工程，並說明其作為生成式AI主要程式介面的角色。
- 区分基礎型LLM與指令微調型LLM，並說明它們如何處理分詞。
- 使用指示、主要內容、提示與範本建構複雜提示。
- 使用
openai庫建構並設定文字生成應用程式，管理環境變數，並透過調整溫度來控制輸出多樣性。 - 区分基於規則的聊天機器人與具上下文感知能力的生成式AI應用程式，同時實踐微軟的「負責人工智能六原則」。
- 透過將文字轉換為嵌入向量（embeddings）並運用餘弦相似度，執行語意搜尋，找出超越簡單關鍵字匹配的相關內容。
- 建構並設定影像生成應用程式，同時實踐「元提示（meta prompts）」以定義內容邊界與安全性。
課程 共 7 课时 · 预计 21.0h
課程
Lesson
This lesson explores the evolution of AI from rigid, rule-based systems to modern Large Language Models that utilize tokenization, attention mechanisms, and statistical probability to generate human-like text. Students will learn how these non-deterministic models function and how to evaluate strategies like prompting, fine-tuning, and RAG to optimize AI performance for specific business needs.
This lesson introduces prompt engineering as the essential interface for guiding generative AI, distinguishing between base models and instruction-tuned models. Students learn to construct effective prompts by mastering tokenization, parameter adjustment, and structural techniques like delimiting to improve model accuracy and reliability.
This lesson explores the transition from traditional, deterministic programming to flexible Generative AI applications by mastering API security, environment configuration, and output control via temperature settings. Students will also learn to implement semantic search using embeddings and apply Microsoft’s Six Principles of Responsible AI to build reliable, context-aware solutions.
This lesson explores how to build secure, integrated AI solutions using low-code platforms like Microsoft Power Platform and advanced techniques such as meta-prompting and function calling. Students learn to bridge the gap between generative models and external data while implementing safety guardrails and deterministic configurations for enterprise-grade applications.
This lesson explores how to build trustworthy Generative AI by balancing user experience pillars, such as explainability and instructional friction, with robust security measures against threats like data poisoning and prompt injection. It also introduces the LLMOps lifecycle, covering the essential stages of ideating, building, and operationalizing AI applications to ensure reliability and performance.
This lesson explores how Retrieval-Augmented Generation (RAG) overcomes LLM knowledge cutoffs and hallucinations by grounding model responses in real-time, external data. Students will learn the technical RAG workflow—including chunking, embeddings, and vector search—while evaluating open-source models and AI agent frameworks for practical implementation.
This lesson explores the optimization hierarchy for generative AI, emphasizing that developers should prioritize prompt engineering and Retrieval-Augmented Generation (RAG) before considering resource-intensive fine-tuning. It also introduces specialized architectures like Small Language Models (SLMs) and Mixture of Experts (MoE) to help balance model performance, inference speed, and deployment efficiency.