@xdev-asia/xdev-knowledge-mcp 1.0.41 → 1.0.42

package/content/series/luyen-thi/luyen-thi-aws-ai-practitioner/chapters/03-domain-3-applications-foundation-models/lessons/07-bai-7-fine-tuning-model-customization.md

@@ -0,0 +1,247 @@
+---
+id: 019c9619-lt01-d3-l07
+title: 'Bài 7: Fine-tuning & Model Customization'
+slug: bai-7-fine-tuning-model-customization
+description: >-
+  Pre-training vs Fine-tuning vs RLHF. PEFT & LoRA.
+  Continued Pre-training. Amazon Bedrock Custom Models.
+  Training data preparation, evaluation, deployment.
+duration_minutes: 50
+is_free: true
+video_url: null
+sort_order: 3
+section_title: "Domain 3: Applications of Foundation Models (28%)"
+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-bai7-finetuning-spectrum.png" alt="Model Customization Spectrum" style="max-width: 800px; width: 100%; border-radius: 12px;" />
+<p><em>Model Customization Spectrum: từ Prompt Engineering đến Pre-training từ đầu</em></p>
+</div>
+
+<h2 id="customization-spectrum"><strong>1. Model Customization Spectrum</strong></h2>
+
+<p>Có nhiều cách để customize FM behavior, từ đơn giản đến phức tạp:</p>
+
+<pre><code class="language-text">Least Effort                                    Most Effort
+──────────────────────────────────────────────────────────
+Prompt       Few-shot      RAG       Fine-      Continued    Pre-
+Engineering  Prompting               tuning     Pre-training training
+──────────────────────────────────────────────────────────
+No training                   ←                 →    Full training
+$ cheapest                    ←                 →    $$$$ most expensive
+Minutes                       ←                 →    Weeks/Months
+</code></pre>
+
+<h2 id="fine-tuning"><strong>2. Fine-tuning</strong></h2>
+
+<p><strong>Fine-tuning</strong> = further training an existing FM on <strong>your specific dataset</strong> to improve performance on your domain/task.</p>
+
+<h3 id="when-fine-tune"><strong>2.1. When to Fine-tune?</strong></h3>
+
+<table>
+<thead><tr><th>Fine-tune When...</th><th>DON'T Fine-tune When...</th></tr></thead>
+<tbody>
+<tr><td>Need specific style, tone, or format</td><td>Just need factual Q&A (use RAG)</td></tr>
+<tr><td>Domain-specific language patterns</td><td>Task works well with prompting</td></tr>
+<tr><td>Improve accuracy on specific tasks</td><td>Don't have labeled training data</td></tr>
+<tr><td>Reduce prompt size (internalize instructions)</td><td>Data changes frequently (use RAG)</td></tr>
+<tr><td>Need consistent output format</td><td>Budget is limited</td></tr>
+</tbody>
+</table>
+
+<h3 id="fine-tune-types"><strong>2.2. Types of Fine-tuning</strong></h3>
+
+<table>
+<thead><tr><th>Type</th><th>What</th><th>Data Format</th><th>Use Case</th></tr></thead>
+<tbody>
+<tr><td><strong>Instruction fine-tuning</strong></td><td>Train on prompt-response pairs</td><td>{"prompt": "...", "completion": "..."}</td><td>Follow instructions better</td></tr>
+<tr><td><strong>Domain adaptation</strong></td><td>Train on domain text</td><td>Domain documents (medical, legal)</td><td>Learn domain terminology</td></tr>
+<tr><td><strong>Task-specific</strong></td><td>Train on specific task examples</td><td>Task input-output pairs</td><td>Classification, extraction</td></tr>
+</tbody>
+</table>
+
+<h2 id="peft"><strong>3. PEFT & LoRA</strong></h2>
+
+<h3 id="peft-overview"><strong>3.1. Parameter-Efficient Fine-Tuning (PEFT)</strong></h3>
+
+<p>Full fine-tuning updates <strong>ALL model parameters</strong> — expensive and needs lots of GPU memory. PEFT methods update only a <strong>small subset of parameters</strong>.</p>
+
+<pre><code class="language-text">Full Fine-tuning:
+  Model: 7 billion parameters
+  Updated: 7 billion parameters (100%)
+  GPU Memory: Very high
+  Cost: $$$$
+
+PEFT (LoRA):
+  Model: 7 billion parameters
+  Updated: ~10 million parameters (0.1%)
+  GPU Memory: Much lower
+  Cost: $$
+</code></pre>
+
+<h3 id="lora"><strong>3.2. LoRA (Low-Rank Adaptation)</strong></h3>
+
+<p>LoRA thêm <strong>small trainable matrices</strong> vào model layers thay vì update toàn bộ weights:</p>
+
+<ul>
+<li>Freezes original model weights</li>
+<li>Adds small "adapter" matrices (rank decomposition)</li>
+<li>Only trains these small adapters</li>
+<li>At inference: merge adapters with original weights</li>
+</ul>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "Which technique reduces the cost of fine-tuning while maintaining quality?" → <strong>LoRA / PEFT</strong>. Key concept: train a small percentage of parameters instead of all.</p>
+</blockquote>
+
+<h2 id="continued-pretraining"><strong>4. Continued Pre-training</strong></h2>
+
+<p><strong>Continued Pre-training</strong> trains the FM on <strong>large amounts of unlabeled domain data</strong> — teaching the model new vocabulary and concepts <em>before</em> fine-tuning on task-specific data.</p>
+
+<pre><code class="language-text">Workflow:
+Base FM → Continued Pre-training → Fine-tuning → Evaluation
+           (domain corpus,           (labeled        (test on
+            unlabeled)                task data)       holdout)
+
+Example:
+Base Claude → Train on 100K medical papers → Fine-tune on
+              (continued pre-training)       medical Q&A pairs
+              Learns: medical terminology,   Learns: how to
+              drug names, procedures         answer clinical questions
+</code></pre>
+
+<h3 id="cpt-vs-ft"><strong>Continued Pre-training vs Fine-tuning:</strong></h3>
+
+<table>
+<thead><tr><th>Aspect</th><th>Continued Pre-training</th><th>Fine-tuning</th></tr></thead>
+<tbody>
+<tr><td><strong>Data</strong></td><td>Large, unlabeled domain text</td><td>Smaller, labeled task data</td></tr>
+<tr><td><strong>Goal</strong></td><td>Learn domain knowledge</td><td>Learn task-specific behavior</td></tr>
+<tr><td><strong>Cost</strong></td><td>More expensive (larger data)</td><td>Less expensive</td></tr>
+<tr><td><strong>When</strong></td><td>Model lacks domain vocabulary</td><td>Model needs to do specific tasks</td></tr>
+</tbody>
+</table>
+
+<h2 id="rlhf"><strong>5. RLHF (Reinforcement Learning from Human Feedback)</strong></h2>
+
+<p>RLHF is used to <strong>align</strong> model outputs with human preferences — making outputs more helpful, truthful, and harmless.</p>
+
+<pre><code class="language-text">RLHF Pipeline:
+1. Collect human feedback    2. Train reward model    3. Optimize with RL
+   "Which response is           Learns: what humans      FM generates →
+    better? A or B?"             prefer                   reward model scores →
+                                                          update FM weights
+</code></pre>
+
+<p>RLHF is mainly done by <strong>FM providers</strong> (Anthropic, Meta, Amazon) — not typically by end users. But you should know the concept for the exam.</p>
+
+<h2 id="bedrock-custom"><strong>6. Amazon Bedrock Custom Models</strong></h2>
+
+<p>Bedrock offers two customization approaches:</p>
+
+<h3 id="bedrock-ft"><strong>6.1. Fine-tuning in Bedrock</strong></h3>
+
+<table>
+<thead><tr><th>Feature</th><th>Detail</th></tr></thead>
+<tbody>
+<tr><td><strong>Supported models</strong></td><td>Amazon Titan, Meta Llama, Cohere</td></tr>
+<tr><td><strong>Data format</strong></td><td>JSONL with prompt-completion pairs</td></tr>
+<tr><td><strong>Data location</strong></td><td>Amazon S3</td></tr>
+<tr><td><strong>Output</strong></td><td>Custom model version in Bedrock</td></tr>
+<tr><td><strong>Provisioned Throughput</strong></td><td>Required to use fine-tuned model</td></tr>
+</tbody>
+</table>
+
+<h3 id="bedrock-cpt"><strong>6.2. Continued Pre-training in Bedrock</strong></h3>
+
+<table>
+<thead><tr><th>Feature</th><th>Detail</th></tr></thead>
+<tbody>
+<tr><td><strong>Supported models</strong></td><td>Amazon Titan, Meta Llama, Cohere</td></tr>
+<tr><td><strong>Data format</strong></td><td>Plain text files (unlabeled)</td></tr>
+<tr><td><strong>Use case</strong></td><td>Domain adaptation before fine-tuning</td></tr>
+</tbody>
+</table>
+
+<h3 id="bedrock-training-data"><strong>6.3. Training Data Preparation</strong></h3>
+
+<pre><code class="language-json">// Fine-tuning data format (JSONL):
+{"prompt": "What is the recommended dosage of Drug X?", "completion": "The recommended dosage of Drug X is 500mg twice daily for adults."}
+{"prompt": "List side effects of Drug X.", "completion": "Common side effects include headache, nausea, and dizziness."}
+</code></pre>
+
+<h3 id="bedrock-model-eval"><strong>6.4. Model Evaluation in Bedrock</strong></h3>
+
+<p>Amazon Bedrock Model Evaluation cho phép so sánh models:</p>
+
+<ul>
+<li><strong>Automatic evaluation</strong>: Built-in metrics (accuracy, robustness, toxicity)</li>
+<li><strong>Human evaluation</strong>: Human reviewers rate model outputs</li>
+<li><strong>Compare models</strong>: Side-by-side comparison of different FMs</li>
+</ul>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "How to compare the quality of two foundation models for a specific use case?" → <strong>Amazon Bedrock Model Evaluation</strong>. Supports both automatic metrics and human evaluation.</p>
+</blockquote>
+
+<h2 id="data-prep"><strong>7. Training Data Best Practices</strong></h2>
+
+<table>
+<thead><tr><th>Practice</th><th>Why</th></tr></thead>
+<tbody>
+<tr><td><strong>High-quality data</strong></td><td>Garbage in = garbage out</td></tr>
+<tr><td><strong>Diverse examples</strong></td><td>Prevent overfitting to narrow patterns</td></tr>
+<tr><td><strong>Balanced classes</strong></td><td>Avoid bias toward majority class</td></tr>
+<tr><td><strong>Clean data</strong></td><td>Remove duplicates, errors, PII</td></tr>
+<tr><td><strong>Sufficient quantity</strong></td><td>Typically 1000+ for fine-tuning</td></tr>
+<tr><td><strong>Train/validation split</strong></td><td>Evaluate on unseen data</td></tr>
+<tr><td><strong>Format consistency</strong></td><td>Same structure for all examples</td></tr>
+</tbody>
+</table>
+
+<h2 id="summary-table"><strong>8. Summary: When to Use What</strong></h2>
+
+<table>
+<thead><tr><th>Scenario</th><th>Best Approach</th></tr></thead>
+<tbody>
+<tr><td>Simple task, model already good at it</td><td>Prompt Engineering</td></tr>
+<tr><td>Need model to follow a specific pattern</td><td>Few-shot Prompting</td></tr>
+<tr><td>Need answers from company documents</td><td>RAG</td></tr>
+<tr><td>Need specific style/tone/format</td><td>Fine-tuning</td></tr>
+<tr><td>Model doesn't know domain vocabulary</td><td>Continued Pre-training + Fine-tuning</td></tr>
+<tr><td>Align with human preferences</td><td>RLHF (done by FM providers)</td></tr>
+</tbody>
+</table>
+
+<h2 id="practice-questions"><strong>9. Practice Questions</strong></h2>
+
+<p><strong>Q1:</strong> A legal firm wants their AI assistant to generate legal documents in a specific firm-approved writing style. They have 5,000 examples of approved documents. Which customization approach is MOST appropriate?</p>
+<ul>
+<li>A) RAG with a knowledge base</li>
+<li>B) Zero-shot prompting</li>
+<li>C) Fine-tuning on the approved document examples ✓</li>
+<li>D) Continued pre-training on legal textbooks</li>
+</ul>
+<p><em>Explanation: Fine-tuning is ideal for teaching a model a specific writing style with labeled examples. RAG is for retrieving information, not learning styles. Continued pre-training would teach legal concepts but not the firm's specific style.</em></p>
+
+<p><strong>Q2:</strong> Which technique allows fine-tuning a large language model while updating only a small fraction of the model's parameters?</p>
+<ul>
+<li>A) Full fine-tuning</li>
+<li>B) LoRA (Low-Rank Adaptation) ✓</li>
+<li>C) Continued pre-training</li>
+<li>D) RLHF</li>
+</ul>
+<p><em>Explanation: LoRA is a PEFT (Parameter-Efficient Fine-Tuning) method that adds small trainable adapter matrices while freezing the original model weights — typically updating less than 1% of total parameters.</em></p>
+
+<p><strong>Q3:</strong> A company fine-tuned a foundation model, but the model performs well on training data and poorly on new data. What is this problem called?</p>
+<ul>
+<li>A) Underfitting</li>
+<li>B) Overfitting ✓</li>
+<li>C) High bias</li>
+<li>D) Data drift</li>
+</ul>
+<p><em>Explanation: Overfitting occurs when a model memorizes training data instead of learning general patterns. Solutions include: more training data, regularization, lower learning rate, early stopping, or data augmentation.</em></p>

package/content/series/luyen-thi/luyen-thi-aws-ai-practitioner/chapters/03-domain-3-applications-foundation-models/lessons/08-bai-8-amazon-bedrock-deep-dive.md

@@ -0,0 +1,276 @@
+---
+id: 019c9619-lt01-d3-l08
+title: 'Bài 8: Amazon Bedrock Deep Dive'
+slug: bai-8-amazon-bedrock-deep-dive
+description: >-
+  Amazon Bedrock: all features. Agents, Guardrails, Model Evaluation.
+  PartyRock playground. Amazon Q Developer & Amazon Q Business.
+  Choosing the right FM. Pricing models.
+duration_minutes: 65
+is_free: true
+video_url: null
+sort_order: 4
+section_title: "Domain 3: Applications of Foundation Models (28%)"
+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-bai8-bedrock-architecture.png" alt="Amazon Bedrock Architecture" style="max-width: 800px; width: 100%; border-radius: 12px;" />
+<p><em>Amazon Bedrock Architecture — Foundation Models, Agents, Guardrails và Knowledge Bases</em></p>
+</div>
+
+<h2 id="bedrock-overview"><strong>1. Amazon Bedrock Overview</strong></h2>
+
+<p><strong>Amazon Bedrock</strong> là fully managed service cung cấp access đến <strong>FMs từ nhiều providers</strong> qua single API, kèm theo tools để customize, deploy, và secure AI applications.</p>
+
+<h3 id="bedrock-key-features"><strong>1.1. Key Value Propositions</strong></h3>
+
+<ul>
+<li><strong>Choice</strong>: Access FMs từ Amazon, Anthropic, Meta, Mistral, Cohere, Stability AI, AI21 Labs</li>
+<li><strong>Customization</strong>: Fine-tuning, continued pre-training, RAG (Knowledge Bases)</li>
+<li><strong>Security</strong>: Data stays in your AWS account, encrypted, not used to train models</li>
+<li><strong>Serverless</strong>: No infrastructure to manage</li>
+<li><strong>Integration</strong>: Native AWS service integration (IAM, CloudWatch, CloudTrail)</li>
+</ul>
+
+<h3 id="fm-providers"><strong>1.2. Foundation Model Providers on Bedrock</strong></h3>
+
+<table>
+<thead><tr><th>Provider</th><th>Models</th><th>Strengths</th></tr></thead>
+<tbody>
+<tr><td><strong>Amazon</strong></td><td>Titan Text, Titan Embeddings, Titan Image Generator</td><td>General purpose, embeddings, image gen</td></tr>
+<tr><td><strong>Anthropic</strong></td><td>Claude 3 Haiku, Sonnet, Opus</td><td>Complex reasoning, analysis, vision</td></tr>
+<tr><td><strong>Meta</strong></td><td>Llama 2, Llama 3</td><td>Open-source, customizable</td></tr>
+<tr><td><strong>Mistral AI</strong></td><td>Mistral, Mixtral</td><td>Fast, efficient, multilingual</td></tr>
+<tr><td><strong>Cohere</strong></td><td>Command, Embed</td><td>Enterprise text, multilingual embeddings</td></tr>
+<tr><td><strong>Stability AI</strong></td><td>Stable Diffusion XL</td><td>Image generation</td></tr>
+<tr><td><strong>AI21 Labs</strong></td><td>Jurassic</td><td>Text generation, summarization</td></tr>
+</tbody>
+</table>
+
+<h2 id="bedrock-features"><strong>2. Bedrock Features Deep Dive</strong></h2>
+
+<h3 id="bedrock-agents"><strong>2.1. Amazon Bedrock Agents</strong></h3>
+
+<p>Agents cho phép FMs <strong>thực hiện multi-step tasks</strong> bằng cách tự động plan, execute actions, và use tools.</p>
+
+<pre><code class="language-text">User: "Book a flight from Hanoi to Tokyo for next Friday"
+
+Agent workflow:
+1. PLAN: Need to search flights, check availability, book
+2. ACTION: Call flight search API → find available flights
+3. OBSERVE: Found 3 flights, cheapest is $450
+4. ACTION: Call booking API → reserve the flight
+5. RESPOND: "Booked VN flight HAN→NRT, Dec 20, $450"
+</code></pre>
+
+<h3 id="agent-components"><strong>Agent Components:</strong></h3>
+
+<table>
+<thead><tr><th>Component</th><th>Purpose</th></tr></thead>
+<tbody>
+<tr><td><strong>Foundation Model</strong></td><td>Brain that reasons and plans</td></tr>
+<tr><td><strong>Instructions</strong></td><td>System prompt defining agent's role</td></tr>
+<tr><td><strong>Action Groups</strong></td><td>APIs the agent can call (Lambda functions or OpenAPI schemas)</td></tr>
+<tr><td><strong>Knowledge Bases</strong></td><td>RAG data sources for information retrieval</td></tr>
+<tr><td><strong>Guardrails</strong></td><td>Safety and compliance filters</td></tr>
+</tbody>
+</table>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "An AI assistant needs to look up order status, check inventory, and process returns" → <strong>Bedrock Agent</strong> with action groups connected to business APIs.</p>
+</blockquote>
+
+<h3 id="bedrock-guardrails"><strong>2.2. Amazon Bedrock Guardrails</strong></h3>
+
+<p>Guardrails implement <strong>safety controls</strong> for AI applications:</p>
+
+<table>
+<thead><tr><th>Guardrail Type</th><th>What it does</th><th>Example</th></tr></thead>
+<tbody>
+<tr><td><strong>Content filters</strong></td><td>Block harmful content categories</td><td>Hate, violence, sexual, insults</td></tr>
+<tr><td><strong>Denied topics</strong></td><td>Block specific topics</td><td>"Don't discuss competitor products"</td></tr>
+<tr><td><strong>Word filters</strong></td><td>Block specific words/phrases</td><td>Profanity, banned terms</td></tr>
+<tr><td><strong>PII filters</strong></td><td>Detect and redact PII</td><td>SSN, credit card numbers, emails</td></tr>
+<tr><td><strong>Contextual grounding</strong></td><td>Check if response is grounded in context</td><td>Prevent hallucination in RAG</td></tr>
+</tbody>
+</table>
+
+<pre><code class="language-text">Guardrails Flow:
+User Input → [Input Guardrails] → FM Processing → [Output Guardrails] → User
+              Check for:                          Check for:
+              - Denied topics                     - Harmful content
+              - Harmful input                     - PII in response
+              - PII in input                      - Off-topic responses
+                                                  - Grounding check
+</code></pre>
+
+<h3 id="bedrock-eval"><strong>2.3. Model Evaluation</strong></h3>
+
+<p>Compare and evaluate FMs for your specific use case:</p>
+
+<ul>
+<li><strong>Automatic evaluation</strong>: BERTScore, accuracy, toxicity metrics</li>
+<li><strong>Human evaluation</strong>: Custom criteria rated by human reviewers</li>
+<li><strong>A/B comparison</strong>: Side-by-side model comparison</li>
+<li><strong>Custom tasks</strong>: Upload your own test dataset</li>
+</ul>
+
+<h3 id="bedrock-playground"><strong>2.4. Bedrock Playgrounds</strong></h3>
+
+<table>
+<thead><tr><th>Playground</th><th>Use Case</th></tr></thead>
+<tbody>
+<tr><td><strong>Text playground</strong></td><td>Test text models interactively</td></tr>
+<tr><td><strong>Chat playground</strong></td><td>Test conversational models</td></tr>
+<tr><td><strong>Image playground</strong></td><td>Test image generation models</td></tr>
+</tbody>
+</table>
+
+<h2 id="partyrock"><strong>3. Amazon PartyRock</strong></h2>
+
+<p><strong>PartyRock</strong> là <strong>free, no-code playground</strong> cho Bedrock — cho phép bất kỳ ai tạo GenAI apps mà không cần AWS account hay coding skills.</p>
+
+<table>
+<thead><tr><th>Feature</th><th>Detail</th></tr></thead>
+<tbody>
+<tr><td><strong>No AWS account needed</strong></td><td>Free to use with social login</td></tr>
+<tr><td><strong>No coding</strong></td><td>Drag-and-drop app builder</td></tr>
+<tr><td><strong>Shareable</strong></td><td>Share apps via URL</td></tr>
+<tr><td><strong>Use case</strong></td><td>Learning, prototyping, experimentation</td></tr>
+</tbody>
+</table>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "A non-technical marketing team wants to experiment with generative AI without an AWS account" → <strong>PartyRock</strong>.</p>
+</blockquote>
+
+<h2 id="amazon-q"><strong>4. Amazon Q</strong></h2>
+
+<h3 id="q-developer"><strong>4.1. Amazon Q Developer</strong></h3>
+
+<p>AI coding assistant cho developers:</p>
+
+<ul>
+<li><strong>Code generation</strong>: Write code from natural language</li>
+<li><strong>Code explanation</strong>: Explain existing code</li>
+<li><strong>Code transformation</strong>: Upgrade Java versions, .NET migrations</li>
+<li><strong>Debugging</strong>: Identify and fix bugs</li>
+<li><strong>Security scanning</strong>: Find vulnerabilities in code</li>
+<li><strong>IDE integration</strong>: VS Code, JetBrains, AWS Console</li>
+</ul>
+
+<h3 id="q-business"><strong>4.2. Amazon Q Business</strong></h3>
+
+<p>AI assistant for business users:</p>
+
+<ul>
+<li><strong>Connect enterprise data</strong>: S3, SharePoint, Confluence, Salesforce, etc.</li>
+<li><strong>Q&A on company data</strong>: Answers based on connected data sources</li>
+<li><strong>Respects access controls</strong>: ACLs from connected systems</li>
+<li><strong>Plugins</strong>: Create tickets (Jira), send emails, etc.</li>
+</ul>
+
+<h3 id="q-vs-bedrock"><strong>4.3. Amazon Q vs Bedrock</strong></h3>
+
+<table>
+<thead><tr><th>Feature</th><th>Amazon Q</th><th>Amazon Bedrock</th></tr></thead>
+<tbody>
+<tr><td><strong>Target user</strong></td><td>End users (devs, business)</td><td>Developers building AI apps</td></tr>
+<tr><td><strong>Customization</strong></td><td>Limited (connect data sources)</td><td>Full (fine-tune, RAG, agents)</td></tr>
+<tr><td><strong>Managed</strong></td><td>Fully managed assistant</td><td>API/SDK access to FMs</td></tr>
+<tr><td><strong>Use case</strong></td><td>Productivity tool</td><td>Building custom AI applications</td></tr>
+</tbody>
+</table>
+
+<h2 id="pricing"><strong>5. Bedrock Pricing Models</strong></h2>
+
+<table>
+<thead><tr><th>Pricing Model</th><th>How it works</th><th>Best For</th></tr></thead>
+<tbody>
+<tr><td><strong>On-Demand</strong></td><td>Pay per input/output token</td><td>Variable, unpredictable workloads</td></tr>
+<tr><td><strong>Provisioned Throughput</strong></td><td>Reserved model units (hourly)</td><td>Consistent, production workloads</td></tr>
+<tr><td><strong>Batch Inference</strong></td><td>Submit batch jobs (up to 50% cheaper)</td><td>Large-scale, non-real-time processing</td></tr>
+</tbody>
+</table>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "Cost-optimize a GenAI workload with predictable traffic?" → <strong>Provisioned Throughput</strong>. "Process thousands of documents overnight?" → <strong>Batch Inference</strong>.</p>
+</blockquote>
+
+<h2 id="choosing-fm"><strong>6. How to Choose the Right FM</strong></h2>
+
+<pre><code class="language-text">Decision Framework:
+┌─────────────────────────────────────────────────┐
+│ 1. TASK TYPE                                    │
+│    Text? Image? Code? Multi-modal?              │
+├─────────────────────────────────────────────────┤
+│ 2. COMPLEXITY                                   │
+│    Simple classification → smaller model         │
+│    Complex reasoning → larger model              │
+├─────────────────────────────────────────────────┤
+│ 3. LATENCY REQUIREMENTS                         │
+│    Real-time → smaller/faster model (Haiku)      │
+│    Batch processing → larger model (Opus)        │
+├─────────────────────────────────────────────────┤
+│ 4. COST CONSTRAINTS                             │
+│    Budget limited → smaller model                │
+│    Quality critical → larger model               │
+├─────────────────────────────────────────────────┤
+│ 5. CUSTOMIZATION NEEDS                          │
+│    Fine-tuning needed? Check supported models    │
+│    LoRA? Check compatibility                     │
+├─────────────────────────────────────────────────┤
+│ 6. EVALUATE with Model Evaluation               │
+│    Test candidates side-by-side                  │
+└─────────────────────────────────────────────────┘
+</code></pre>
+
+<h2 id="other-services"><strong>7. Other AWS GenAI Services</strong></h2>
+
+<table>
+<thead><tr><th>Service</th><th>What it does</th></tr></thead>
+<tbody>
+<tr><td><strong>Amazon CodeWhisperer</strong></td><td>Now part of Amazon Q Developer (code suggestions)</td></tr>
+<tr><td><strong>AWS App Studio</strong></td><td>Build enterprise apps with natural language</td></tr>
+<tr><td><strong>Amazon SageMaker JumpStart</strong></td><td>Deploy open-source FMs with SageMaker</td></tr>
+<tr><td><strong>Amazon Comprehend</strong></td><td>NLP service (sentiment, entities, topics — pre-built)</td></tr>
+<tr><td><strong>Amazon Transcribe</strong></td><td>Speech-to-text</td></tr>
+<tr><td><strong>Amazon Polly</strong></td><td>Text-to-speech</td></tr>
+<tr><td><strong>Amazon Translate</strong></td><td>Machine translation</td></tr>
+<tr><td><strong>Amazon Rekognition</strong></td><td>Image/video analysis</td></tr>
+<tr><td><strong>Amazon Textract</strong></td><td>Extract text from documents (OCR+)</td></tr>
+</tbody>
+</table>
+
+<h2 id="practice-questions"><strong>8. Practice Questions</strong></h2>
+
+<p><strong>Q1:</strong> A retail company wants to build an AI assistant that can check inventory, process returns, and answer product questions from their catalog. Which Amazon Bedrock feature should they use?</p>
+<ul>
+<li>A) Bedrock Guardrails</li>
+<li>B) Bedrock Knowledge Bases only</li>
+<li>C) Bedrock Agents with Action Groups and Knowledge Bases ✓</li>
+<li>D) Bedrock Model Evaluation</li>
+</ul>
+<p><em>Explanation: Bedrock Agents can orchestrate multi-step tasks by calling APIs (action groups for inventory/returns) and retrieving information (knowledge bases for product catalog).</em></p>
+
+<p><strong>Q2:</strong> Which Amazon Bedrock feature should be used to prevent a generative AI application from discussing competitor products and to filter out personally identifiable information (PII)?</p>
+<ul>
+<li>A) Bedrock Knowledge Bases</li>
+<li>B) Bedrock Custom Models</li>
+<li>C) Bedrock Guardrails ✓</li>
+<li>D) Bedrock Agents</li>
+</ul>
+<p><em>Explanation: Guardrails provide denied topic filtering (block competitor discussions) and PII detection/redaction. They can be applied to both input and output of FM calls.</em></p>
+
+<p><strong>Q3:</strong> A company wants to process 50,000 customer reviews overnight for sentiment analysis using a foundation model. Which Bedrock pricing model is MOST cost-effective?</p>
+<ul>
+<li>A) On-Demand pricing</li>
+<li>B) Provisioned Throughput</li>
+<li>C) Batch Inference ✓</li>
+<li>D) Free tier</li>
+</ul>
+<p><em>Explanation: Batch Inference is designed for large-scale, non-real-time workloads and offers up to 50% cost savings compared to on-demand pricing. Ideal for overnight processing.</em></p>