@xdev-asia/xdev-knowledge-mcp 1.0.40 → 1.0.42

package/content/series/luyen-thi/luyen-thi-aws-ai-practitioner/chapters/02-domain-2-fundamentals-generative-ai/lessons/03-bai-3-generative-ai-foundation-models.md

@@ -0,0 +1,218 @@
+---
+id: 019c9619-lt01-d2-l03
+title: 'Bài 3: Generative AI & Foundation Models'
+slug: bai-3-generative-ai-foundation-models
+description: >-
+  Generative AI là gì. Foundation Models: pre-training, fine-tuning.
+  Types: text-to-text, text-to-image, text-to-code. Tokenization.
+  Model parameters, inference, temperature, top-p, top-k.
+duration_minutes: 60
+is_free: true
+video_url: null
+sort_order: 1
+section_title: "Domain 2: Fundamentals of Generative AI (24%)"
+course:
+  id: 019c9619-lt01-7001-c001-lt0100000001
+  title: 'Luyện thi AWS Certified AI Practitioner (AIF-C01)'
+  slug: luyen-thi-aws-ai-practitioner
+---
+
+<div style="text-align: center; margin: 2rem 0;">
+<img src="/storage/uploads/2026/04/aws-aif-bai3-foundation-model-lifecycle.png" alt="Foundation Model Lifecycle" style="max-width: 800px; width: 100%; border-radius: 12px;" />
+<p><em>Foundation Model Lifecycle — Pre-training, Fine-tuning, RAG và Prompt Engineering</em></p>
+</div>
+
+<h2 id="overview"><strong>Tổng quan Domain 2</strong></h2>
+
+<p>Domain 2 chiếm <strong>24% đề thi</strong> — đây là domain lớn thứ hai. Bạn cần hiểu rõ Generative AI, Foundation Models, và cách chúng khác biệt so với traditional ML.</p>
+
+<h2 id="what-is-genai"><strong>1. What is Generative AI?</strong></h2>
+
+<p><strong>Generative AI</strong> là nhánh của AI tập trung vào việc <strong>tạo nội dung mới</strong> (text, images, code, audio, video) dựa trên patterns học được từ training data.</p>
+
+<h3 id="discriminative-vs-generative"><strong>Discriminative vs Generative AI</strong></h3>
+
+<table>
+<thead><tr><th>Aspect</th><th>Discriminative AI</th><th>Generative AI</th></tr></thead>
+<tbody>
+<tr><td><strong>What it does</strong></td><td>Classify / predict</td><td>Create / generate</td></tr>
+<tr><td><strong>Output</strong></td><td>Label, category, number</td><td>New content (text, image, code)</td></tr>
+<tr><td><strong>Example</strong></td><td>"Is this email spam?" → Yes/No</td><td>"Write an email about..." → New email</td></tr>
+<tr><td><strong>Models</strong></td><td>Logistic Regression, SVM, CNN classifier</td><td>GPT, Claude, Stable Diffusion, DALL-E</td></tr>
+</tbody>
+</table>
+
+<h3 id="genai-modalities"><strong>Generative AI Modalities</strong></h3>
+
+<table>
+<thead><tr><th>Input → Output</th><th>Examples</th><th>Models</th></tr></thead>
+<tbody>
+<tr><td><strong>Text → Text</strong></td><td>Chatbot, summarization, translation</td><td>GPT-4, Claude, Llama</td></tr>
+<tr><td><strong>Text → Image</strong></td><td>Image generation from description</td><td>DALL-E, Stable Diffusion, Titan Image Generator</td></tr>
+<tr><td><strong>Text → Code</strong></td><td>Code generation, debugging</td><td>CodeWhisperer, Copilot</td></tr>
+<tr><td><strong>Text → Audio</strong></td><td>Speech synthesis, music generation</td><td>Amazon Polly (TTS)</td></tr>
+<tr><td><strong>Image → Text</strong></td><td>Image captioning, visual Q&A</td><td>Claude (multi-modal), GPT-4V</td></tr>
+<tr><td><strong>Audio → Text</strong></td><td>Transcription</td><td>Amazon Transcribe, Whisper</td></tr>
+</tbody>
+</table>
+
+<h2 id="foundation-models"><strong>2. Foundation Models</strong></h2>
+
+<p><strong>Foundation Model (FM)</strong> là model AI cực lớn, được <strong>pre-trained trên massive datasets</strong>, có thể adapt cho nhiều downstream tasks khác nhau.</p>
+
+<h3 id="fm-characteristics"><strong>Key Characteristics</strong></h3>
+
+<ul>
+<li><strong>Large-scale pre-training</strong>: Trained on billions of data points (text from internet, books, code)</li>
+<li><strong>General-purpose</strong>: Can handle multiple tasks without task-specific training</li>
+<li><strong>Adaptable</strong>: Can be fine-tuned or prompted for specific use cases</li>
+<li><strong>Expensive to train</strong>: Requires massive compute (GPU/TPU clusters)</li>
+<li><strong>Accessible via API</strong>: Users don't need to train — use through APIs (Amazon Bedrock)</li>
+</ul>
+
+<h3 id="fm-lifecycle"><strong>Foundation Model Lifecycle</strong></h3>
+
+<pre><code class="language-text">┌─────────────────┐     ┌──────────────┐     ┌──────────────┐
+│ 1. Pre-training │────→│ 2. Fine-     │────→│ 3. Inference │
+│ (Massive data,  │     │ tuning       │     │ (Use model   │
+│  Billion params,│     │ (Adapt to    │     │  via API or  │
+│  Very expensive)│     │  specific    │     │  endpoint)   │
+│                 │     │  domain)     │     │              │
+└─────────────────┘     └──────────────┘     └──────────────┘
+     Model Provider          You/Org              Users
+   (Anthropic, Meta,                         (Applications)
+    Amazon, etc.)
+</code></pre>
+
+<h2 id="tokenization"><strong>3. Tokenization</strong></h2>
+
+<p><strong>Tokenization</strong> là quá trình chia text thành các đơn vị nhỏ (<strong>tokens</strong>) mà model hiểu được.</p>
+
+<pre><code class="language-text">Input:  "Machine learning is amazing!"
+Tokens: ["Machine", " learning", " is", " amazing", "!"]
+         token_1    token_2      token_3  token_4    token_5
+
+OR (subword tokenization):
+Tokens: ["Mach", "ine", " learn", "ing", " is", " amaz", "ing", "!"]
+</code></pre>
+
+<h3 id="token-key-points"><strong>Key Concepts for Exam:</strong></h3>
+
+<ul>
+<li><strong>Token ≠ word</strong>: A token can be part of a word, a whole word, or punctuation</li>
+<li><strong>Context window</strong>: Maximum number of tokens a model can process at once (input + output)</li>
+<li><strong>Token limit</strong>: Determines how much text the model can "see" and generate</li>
+<li><strong>Pricing</strong>: API calls are typically priced per token (input tokens + output tokens)</li>
+</ul>
+
+<blockquote>
+<p><strong>Exam tip:</strong> Context window size matters. Larger context = can process longer documents. But costs more and may be slower.</p>
+</blockquote>
+
+<h2 id="model-parameters"><strong>4. Model Parameters & Inference Settings</strong></h2>
+
+<h3 id="model-params"><strong>4.1. Model Parameters (Learned during training)</strong></h3>
+
+<ul>
+<li><strong>Parameters</strong> = weights and biases trong neural network</li>
+<li>GPT-4: ~1.7 trillion parameters, Claude: undisclosed, Llama 3: 8B/70B/405B</li>
+<li>More parameters → generally more capable, but more expensive</li>
+</ul>
+
+<h3 id="inference-params"><strong>4.2. Inference Parameters (Set by user)</strong></h3>
+
+<p>Khi gọi model, bạn có thể điều chỉnh các <strong>inference parameters</strong>:</p>
+
+<table>
+<thead><tr><th>Parameter</th><th>Range</th><th>What it controls</th></tr></thead>
+<tbody>
+<tr><td><strong>Temperature</strong></td><td>0.0 → 1.0+</td><td>Randomness/creativity. Low = deterministic, focused. High = creative, diverse.</td></tr>
+<tr><td><strong>Top-p (Nucleus)</strong></td><td>0.0 → 1.0</td><td>Cumulative probability threshold. Lower = more focused vocabulary.</td></tr>
+<tr><td><strong>Top-k</strong></td><td>1 → ∞</td><td>Number of top tokens to consider. Lower = more predictable.</td></tr>
+<tr><td><strong>Max tokens</strong></td><td>1 → limit</td><td>Maximum length of generated output.</td></tr>
+<tr><td><strong>Stop sequences</strong></td><td>strings</td><td>Text that tells model to stop generating.</td></tr>
+</tbody>
+</table>
+
+<h3 id="temperature-guide"><strong>Temperature Guide for Exam</strong></h3>
+
+<pre><code class="language-text">Temperature = 0  →  Most deterministic (factual Q&A, code, data extraction)
+Temperature = 0.3 → Slightly creative (business writing, summaries)
+Temperature = 0.7 → Creative (stories, brainstorming, marketing copy)
+Temperature = 1.0+ → Very random (poetry, creative writing — may hallucinate more)
+</code></pre>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "A company needs consistent, accurate answers for customer FAQ" → use <strong>low temperature</strong>. "A company wants creative marketing slogans" → use <strong>high temperature</strong>.</p>
+</blockquote>
+
+<h2 id="hallucination"><strong>5. Hallucination</strong></h2>
+
+<p><strong>Hallucination</strong> là khi model tạo ra output <strong>confident nhưng incorrect</strong> — bịa ra facts, citations, hoặc thông tin không tồn tại.</p>
+
+<h3 id="hallucination-causes"><strong>Causes:</strong></h3>
+<ul>
+<li>Training data gaps or outdated information</li>
+<li>Model doesn't truly "know" facts — it predicts likely next tokens</li>
+<li>Ambiguous or too-open prompts</li>
+<li>High temperature settings</li>
+</ul>
+
+<h3 id="hallucination-mitigation"><strong>Mitigation Strategies:</strong></h3>
+<table>
+<thead><tr><th>Strategy</th><th>How it helps</th></tr></thead>
+<tbody>
+<tr><td><strong>RAG</strong> (Retrieval-Augmented Generation)</td><td>Ground responses in actual data from knowledge base</td></tr>
+<tr><td><strong>Lower temperature</strong></td><td>Reduce randomness in generation</td></tr>
+<tr><td><strong>Guardrails</strong></td><td>Filter/validate outputs (Amazon Bedrock Guardrails)</td></tr>
+<tr><td><strong>Better prompts</strong></td><td>"Only answer based on provided context" / "Say I don't know if unsure"</td></tr>
+<tr><td><strong>Fine-tuning</strong></td><td>Train model on domain-specific accurate data</td></tr>
+<tr><td><strong>Human review</strong></td><td>Human-in-the-loop validation</td></tr>
+</tbody>
+</table>
+
+<h2 id="fm-on-aws"><strong>6. Foundation Models on AWS (Amazon Bedrock)</strong></h2>
+
+<p>Amazon Bedrock cung cấp access đến nhiều Foundation Models từ các providers:</p>
+
+<table>
+<thead><tr><th>Provider</th><th>Models</th><th>Strengths</th></tr></thead>
+<tbody>
+<tr><td><strong>Anthropic</strong></td><td>Claude 3 (Haiku, Sonnet, Opus)</td><td>Reasoning, safety, long context</td></tr>
+<tr><td><strong>Meta</strong></td><td>Llama 3</td><td>Open-source, versatile</td></tr>
+<tr><td><strong>Amazon</strong></td><td>Titan (Text, Embeddings, Image)</td><td>AWS-native, embeddings for RAG</td></tr>
+<tr><td><strong>Mistral AI</strong></td><td>Mistral, Mixtral</td><td>Efficient, fast inference</td></tr>
+<tr><td><strong>Stability AI</strong></td><td>Stable Diffusion</td><td>Image generation</td></tr>
+<tr><td><strong>Cohere</strong></td><td>Command, Embed</td><td>Enterprise NLP, embeddings</td></tr>
+<tr><td><strong>AI21 Labs</strong></td><td>Jurassic</td><td>Text generation</td></tr>
+</tbody>
+</table>
+
+<h2 id="practice-questions"><strong>7. Practice Questions</strong></h2>
+
+<p><strong>Q1:</strong> What is the PRIMARY advantage of Foundation Models compared to traditional ML models?</p>
+<ul>
+<li>A) They are smaller and faster</li>
+<li>B) They can be adapted to multiple downstream tasks without task-specific training ✓</li>
+<li>C) They never produce incorrect outputs</li>
+<li>D) They don't require any compute resources</li>
+</ul>
+<p><em>Explanation: Foundation Models are pre-trained on massive datasets and can be adapted (via prompting or fine-tuning) for many different tasks. They are large, can hallucinate, and still require compute.</em></p>
+
+<p><strong>Q2:</strong> A company uses a generative AI model and notices it sometimes generates plausible but factually incorrect information. What is this phenomenon called?</p>
+<ul>
+<li>A) Overfitting</li>
+<li>B) Data drift</li>
+<li>C) Hallucination ✓</li>
+<li>D) Bias</li>
+</ul>
+<p><em>Explanation: Hallucination is when a generative AI model produces confident but factually incorrect outputs.</em></p>
+
+<p><strong>Q3:</strong> A developer wants to ensure their generative AI chatbot provides consistent, factual answers with minimal creativity. Which inference parameter should they adjust?</p>
+<ul>
+<li>A) Set max tokens to a very high value</li>
+<li>B) Set temperature close to 0 ✓</li>
+<li>C) Set temperature close to 1</li>
+<li>D) Increase the top-k value</li>
+</ul>
+<p><em>Explanation: Low temperature makes the model more deterministic and focused, reducing creativity and randomness in responses.</em></p>

package/content/series/luyen-thi/luyen-thi-aws-ai-practitioner/chapters/02-domain-2-fundamentals-generative-ai/lessons/04-bai-4-llm-transformers-multimodal.md

@@ -0,0 +1,232 @@
+---
+id: 019c9619-lt01-d2-l04
+title: 'Bài 4: LLMs, Transformers & Multi-modal Models'
+slug: bai-4-llm-transformers-multimodal
+description: >-
+  Transformer architecture: attention mechanism, self-attention.
+  GPT (decoder-only), BERT (encoder-only), T5 (encoder-decoder).
+  Multi-modal models. Hallucination: causes and mitigation.
+  Embeddings và vector representations.
+duration_minutes: 60
+is_free: true
+video_url: null
+sort_order: 2
+section_title: "Domain 2: Fundamentals of Generative AI (24%)"
+course:
+  id: 019c9619-lt01-7001-c001-lt0100000001
+  title: 'Luyện thi AWS Certified AI Practitioner (AIF-C01)'
+  slug: luyen-thi-aws-ai-practitioner
+---
+
+<div style="text-align: center; margin: 2rem 0;">
+<img src="/storage/uploads/2026/04/aws-aif-bai4-transformer-architecture.png" alt="Transformer Architecture" style="max-width: 800px; width: 100%; border-radius: 12px;" />
+<p><em>Transformer Architecture — Encoder stack, Decoder stack và các biến thể BERT/GPT/T5</em></p>
+</div>
+
+<h2 id="transformer"><strong>1. Transformer Architecture</strong></h2>
+
+<p>Transformer là kiến trúc neural network đã <strong>cách mạng hoá NLP</strong>, được giới thiệu trong paper "Attention Is All You Need" (2017). Hầu hết LLMs hiện tại đều dựa trên Transformer.</p>
+
+<h3 id="attention"><strong>1.1. Self-Attention Mechanism</strong></h3>
+
+<p>Self-attention cho phép model xem xét <strong>mối quan hệ giữa tất cả các từ</strong> trong input, bất kể khoảng cách.</p>
+
+<pre><code class="language-text">Input: "The cat sat on the mat because it was tired"
+
+Self-attention answers: What does "it" refer to?
+→ Attends to "cat" (high attention score)
+→ Not "mat" (low attention score)
+
+Traditional RNN would struggle with this long-range dependency.
+</code></pre>
+
+<h3 id="encoder-decoder"><strong>1.2. Encoder-Decoder Architecture</strong></h3>
+
+<pre><code class="language-text">Original Transformer:
+┌──────────────────────────┐
+│        ENCODER           │  ← Understands input
+│  (Self-Attention +       │
+│   Feed-Forward layers)   │
+├──────────────────────────┤
+│        DECODER           │  ← Generates output
+│  (Masked Self-Attention +│
+│   Cross-Attention +      │
+│   Feed-Forward layers)   │
+└──────────────────────────┘
+</code></pre>
+
+<h3 id="transformer-types"><strong>1.3. Three Types of Transformers</strong></h3>
+
+<table>
+<thead><tr><th>Type</th><th>Architecture</th><th>Best For</th><th>Models</th></tr></thead>
+<tbody>
+<tr><td><strong>Encoder-only</strong></td><td>Encoder</td><td>Understanding text (classification, NER, sentiment)</td><td>BERT, RoBERTa, DistilBERT</td></tr>
+<tr><td><strong>Decoder-only</strong></td><td>Decoder</td><td>Generating text (chatbot, content creation)</td><td>GPT-4, Claude, Llama</td></tr>
+<tr><td><strong>Encoder-Decoder</strong></td><td>Both</td><td>Sequence-to-sequence (translation, summarization)</td><td>T5, BART</td></tr>
+</tbody>
+</table>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "Which architecture is best for text generation?" → <strong>Decoder-only</strong> (GPT, Claude). "Which architecture is best for text classification?" → <strong>Encoder-only</strong> (BERT).</p>
+</blockquote>
+
+<h2 id="llm"><strong>2. Large Language Models (LLMs)</strong></h2>
+
+<p>LLMs là Foundation Models specifically для text — trained on massive text corpora to understand and generate human language.</p>
+
+<h3 id="llm-capabilities"><strong>2.1. LLM Capabilities</strong></h3>
+
+<table>
+<thead><tr><th>Capability</th><th>Description</th><th>Example</th></tr></thead>
+<tbody>
+<tr><td><strong>Text Generation</strong></td><td>Create new text content</td><td>Articles, emails, stories</td></tr>
+<tr><td><strong>Summarization</strong></td><td>Condense long text</td><td>Document summaries</td></tr>
+<tr><td><strong>Translation</strong></td><td>Convert between languages</td><td>English → Vietnamese</td></tr>
+<tr><td><strong>Q&A</strong></td><td>Answer questions</td><td>Customer support, FAQ</td></tr>
+<tr><td><strong>Code Generation</strong></td><td>Write and explain code</td><td>Amazon Q Developer</td></tr>
+<tr><td><strong>Text Classification</strong></td><td>Categorize text</td><td>Sentiment analysis</td></tr>
+<tr><td><strong>Reasoning</strong></td><td>Logical analysis</td><td>Math problems, step-by-step reasoning</td></tr>
+</tbody>
+</table>
+
+<h3 id="llm-limitations"><strong>2.2. LLM Limitations</strong></h3>
+
+<ul>
+<li><strong>Knowledge cutoff</strong>: Doesn't know events after training data cutoff date</li>
+<li><strong>Hallucination</strong>: Can generate false information confidently</li>
+<li><strong>Context window limit</strong>: Can't process unlimited text</li>
+<li><strong>No real-time data</strong>: Can't access internet or live data (unless augmented)</li>
+<li><strong>Expensive</strong>: Large models need significant compute for inference</li>
+<li><strong>Bias</strong>: Can reflect biases in training data</li>
+</ul>
+
+<h2 id="embeddings"><strong>3. Embeddings & Vector Representations</strong></h2>
+
+<p><strong>Embeddings</strong> biến text (hoặc images, audio) thành <strong>numerical vectors</strong> mà machines hiểu được. Các text có ý nghĩa tương tự sẽ có vectors gần nhau trong không gian nhiều chiều.</p>
+
+<pre><code class="language-text">Text: "King"     → [0.23, 0.87, -0.12, 0.45, ...]
+Text: "Queen"    → [0.21, 0.89, -0.15, 0.43, ...]  ← Close vectors!
+Text: "Banana"   → [0.91, -0.32, 0.67, -0.88, ...] ← Far away
+
+Relationship: King - Man + Woman ≈ Queen
+</code></pre>
+
+<h3 id="embeddings-use"><strong>Why Embeddings Matter for the Exam:</strong></h3>
+
+<ul>
+<li><strong>Semantic search</strong>: Find similar documents based on meaning (not just keywords)</li>
+<li><strong>RAG</strong>: Convert documents to embeddings, store in vector DB, retrieve relevant context</li>
+<li><strong>Clustering</strong>: Group similar documents/sentences</li>
+<li><strong>Amazon Titan Embeddings</strong>: AWS model specifically for creating text embeddings</li>
+</ul>
+
+<h3 id="vector-db"><strong>Vector Databases</strong></h3>
+
+<p>Store and search embeddings efficiently:</p>
+
+<table>
+<thead><tr><th>Vector DB</th><th>Notes</th></tr></thead>
+<tbody>
+<tr><td><strong>Amazon OpenSearch Serverless</strong></td><td>AWS-managed vector search</td></tr>
+<tr><td><strong>Amazon Aurora (pgvector)</strong></td><td>PostgreSQL with vector extension</td></tr>
+<tr><td><strong>Pinecone</strong></td><td>Popular third-party vector DB</td></tr>
+<tr><td><strong>Amazon Bedrock Knowledge Bases</strong></td><td>Managed RAG — handles vector storage internally</td></tr>
+</tbody>
+</table>
+
+<h2 id="multimodal"><strong>4. Multi-modal Models</strong></h2>
+
+<p><strong>Multi-modal models</strong> có thể xử lý và tạo nội dung từ <strong>nhiều loại data types</strong> (text + images + audio + video).</p>
+
+<h3 id="multimodal-examples"><strong>Examples on AWS:</strong></h3>
+
+<table>
+<thead><tr><th>Model</th><th>Modalities</th><th>What it can do</th></tr></thead>
+<tbody>
+<tr><td><strong>Claude 3</strong> (Anthropic)</td><td>Text + Image input → Text output</td><td>Describe images, analyze charts, visual Q&A</td></tr>
+<tr><td><strong>Amazon Titan Image Generator</strong></td><td>Text → Image</td><td>Create images from text descriptions</td></tr>
+<tr><td><strong>Amazon Titan Multimodal Embeddings</strong></td><td>Text + Image → Vectors</td><td>Search across text and images</td></tr>
+<tr><td><strong>Stable Diffusion</strong> (Stability AI)</td><td>Text → Image</td><td>Generate and edit images</td></tr>
+</tbody>
+</table>
+
+<h3 id="multimodal-usecases"><strong>Multi-modal Use Cases for Exam:</strong></h3>
+
+<ul>
+<li>"Analyze product images and generate descriptions" → Multi-modal model (Claude 3 Vision)</li>
+<li>"Generate product images from text descriptions" → Text-to-image (Titan Image Generator, Stable Diffusion)</li>
+<li>"Search across both text documents and images" → Multi-modal embeddings</li>
+</ul>
+
+<h2 id="diffusion"><strong>5. Diffusion Models</strong></h2>
+
+<p>Diffusion models (như Stable Diffusion) hoạt động bằng cách:</p>
+
+<ol>
+<li><strong>Forward process</strong>: Gradually add noise to an image until it becomes pure noise</li>
+<li><strong>Reverse process</strong>: Learn to remove noise step by step, generating a new image</li>
+</ol>
+
+<pre><code class="language-text">Training (Forward):
+Clean Image → Add Noise → Add More Noise → ... → Pure Noise
+
+Generation (Reverse):
+Pure Noise → Remove Noise → Remove More Noise → ... → New Image
+                           (guided by text prompt)
+</code></pre>
+
+<blockquote>
+<p><strong>Exam tip:</strong> Bạn không cần biết math chi tiết, chỉ cần hiểu concept: diffusion models tạo images bằng cách <strong>từ từ khử noise có guided bởi text prompt</strong>.</p>
+</blockquote>
+
+<h2 id="training-types"><strong>6. Pre-training vs Fine-tuning vs Prompting</strong></h2>
+
+<table>
+<thead><tr><th>Method</th><th>What</th><th>Data Needed</th><th>Cost</th><th>When to Use</th></tr></thead>
+<tbody>
+<tr><td><strong>Pre-training</strong></td><td>Train from scratch</td><td>Billions of examples</td><td>$$$$</td><td>Creating new FM (done by providers)</td></tr>
+<tr><td><strong>Fine-tuning</strong></td><td>Further train existing FM</td><td>Thousands of examples</td><td>$$</td><td>Domain-specific knowledge</td></tr>
+<tr><td><strong>Prompt Engineering</strong></td><td>Craft better inputs</td><td>None (few examples)</td><td>$</td><td>Quick adaptation, no training needed</td></tr>
+<tr><td><strong>RAG</strong></td><td>Augment with external data</td><td>Knowledge base</td><td>$</td><td>Access current/proprietary data</td></tr>
+</tbody>
+</table>
+
+<h3 id="decision-tree"><strong>Decision Tree for Exam:</strong></h3>
+
+<pre><code class="language-text">Need the model to know specific domain knowledge?
+├── Is the knowledge in documents you can provide?
+│   └── YES → RAG (Bedrock Knowledge Bases)
+│   └── NO, model needs to learn patterns →
+│       ├── Have thousands of training examples? → Fine-tuning
+│       └── Only a few examples? → Few-shot prompting
+├── General knowledge is enough? → Prompt Engineering (zero/few-shot)
+</code></pre>
+
+<h2 id="practice-questions"><strong>7. Practice Questions</strong></h2>
+
+<p><strong>Q1:</strong> A company wants to search for relevant information across both product images and text descriptions. Which type of model would be MOST suitable?</p>
+<ul>
+<li>A) A text-only LLM</li>
+<li>B) A multi-modal embedding model ✓</li>
+<li>C) A diffusion model</li>
+<li>D) A RNN model</li>
+</ul>
+<p><em>Explanation: Multi-modal embedding models can create vector representations of both text and images in the same vector space, enabling cross-modal search.</em></p>
+
+<p><strong>Q2:</strong> Which Transformer architecture is BEST suited for text generation tasks such as chatbots and content creation?</p>
+<ul>
+<li>A) Encoder-only (BERT)</li>
+<li>B) Decoder-only (GPT, Claude) ✓</li>
+<li>C) Encoder-decoder (T5)</li>
+<li>D) Convolutional Neural Network (CNN)</li>
+</ul>
+<p><em>Explanation: Decoder-only architectures generate text one token at a time (autoregressive) and are the basis for most modern chatbots and text generators.</em></p>
+
+<p><strong>Q3:</strong> What is the purpose of text embeddings in the context of generative AI applications?</p>
+<ul>
+<li>A) To compress files for storage</li>
+<li>B) To convert text into numerical vectors that capture semantic meaning ✓</li>
+<li>C) To encrypt text for security</li>
+<li>D) To translate text between languages</li>
+</ul>
+<p><em>Explanation: Embeddings are numerical vector representations of text that capture semantic meaning. Similar texts have similar vectors, enabling semantic search, RAG, and clustering.</em></p>