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

package/content/series/luyen-thi/luyen-thi-aws-ai-practitioner/chapters/03-domain-3-applications-foundation-models/lessons/05-bai-5-prompt-engineering-techniques.md

@@ -0,0 +1,254 @@
+---
+id: 019c9619-lt01-d3-l05
+title: 'Bài 5: Prompt Engineering Techniques'
+slug: bai-5-prompt-engineering-techniques
+description: >-
+  Zero-shot, few-shot, Chain-of-Thought prompting.
+  System prompts, prompt templates, negative prompting.
+  Prompt engineering best practices cho AWS AI Practitioner exam.
+duration_minutes: 55
+is_free: true
+video_url: null
+sort_order: 1
+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-bai5-prompt-engineering.png" alt="Prompt Engineering Techniques" style="max-width: 800px; width: 100%; border-radius: 12px;" />
+<p><em>Prompt Engineering Techniques: Zero-shot, Few-shot và Chain-of-Thought</em></p>
+</div>
+
+<h2 id="prompt-engineering"><strong>1. What is Prompt Engineering?</strong></h2>
+
+<p><strong>Prompt Engineering</strong> là nghệ thuật thiết kế input (prompt) để nhận được output mong muốn từ Foundation Models. Đây là cách <strong>rẻ nhất và nhanh nhất</strong> để customize FM behavior — không cần training hay fine-tuning.</p>
+
+<h3 id="prompt-components"><strong>1.1. Components of a Prompt</strong></h3>
+
+<pre><code class="language-text">┌────────────────────────────────────────────┐
+│  SYSTEM PROMPT (optional)                  │
+│  "You are a helpful AWS solutions          │
+│   architect. Answer concisely."            │
+├────────────────────────────────────────────┤
+│  CONTEXT (optional)                        │
+│  Background info, documents, data          │
+├────────────────────────────────────────────┤
+│  USER PROMPT (required)                    │
+│  The actual question or instruction        │
+├────────────────────────────────────────────┤
+│  EXAMPLES (optional, for few-shot)         │
+│  Input → Output pairs                     │
+├────────────────────────────────────────────┤
+│  OUTPUT FORMAT (optional)                  │
+│  "Respond in JSON", "Use bullet points"    │
+└────────────────────────────────────────────┘
+</code></pre>
+
+<h2 id="prompting-techniques"><strong>2. Prompting Techniques</strong></h2>
+
+<h3 id="zero-shot"><strong>2.1. Zero-shot Prompting</strong></h3>
+
+<p>Gửi prompt <strong>không có ví dụ</strong>. Model dựa hoàn toàn vào knowledge đã học.</p>
+
+<pre><code class="language-text">Prompt: "Classify the sentiment of this review:
+'The product arrived damaged and customer service was unhelpful.'
+
+Sentiment:"
+
+Output: "Negative"
+</code></pre>
+
+<p><strong>When to use:</strong> Simple, well-defined tasks mà model đã hiểu rõ.</p>
+
+<h3 id="few-shot"><strong>2.2. Few-shot Prompting</strong></h3>
+
+<p>Cung cấp <strong>một vài examples</strong> trước khi đưa ra task thực tế. Giúp model hiểu expected format và logic.</p>
+
+<pre><code class="language-text">Prompt: "Classify these reviews:
+
+Review: 'Amazing quality, fast shipping!' → Positive
+Review: 'Terrible experience, never again.' → Negative
+Review: 'It's okay, nothing special.' → Neutral
+
+Review: 'The product exceeded my expectations!' →"
+
+Output: "Positive"
+</code></pre>
+
+<p><strong>When to use:</strong> Khi cần model tuân theo specific format hoặc logic pattern mà zero-shot không đạt đủ chất lượng.</p>
+
+<h3 id="one-shot"><strong>2.3. One-shot Prompting</strong></h3>
+
+<p>Variation của few-shot nhưng chỉ cung cấp <strong>1 example</strong>. Dùng khi bạn muốn set pattern nhưng context window bị hạn chế.</p>
+
+<h3 id="cot"><strong>2.4. Chain-of-Thought (CoT) Prompting</strong></h3>
+
+<p>Yêu cầu model <strong>suy nghĩ step-by-step</strong> trước khi trả lời. Đặc biệt hiệu quả cho math, logic, và reasoning tasks.</p>
+
+<pre><code class="language-text">WITHOUT CoT:
+Q: "If a store has 3 boxes with 12 apples each, and gives
+    away 15 apples, how many are left?"
+A: "21" (might be wrong without reasoning)
+
+WITH CoT:
+Q: "Think step by step: If a store has 3 boxes with 12
+    apples each, and gives away 15 apples, how many are left?"
+A: "Step 1: Total apples = 3 × 12 = 36
+    Step 2: After giving away = 36 - 15 = 21
+    Answer: 21 apples"
+</code></pre>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "Which prompting technique improves reasoning accuracy?" → <strong>Chain-of-Thought</strong>. Key phrase: "think step by step" hoặc "explain your reasoning".</p>
+</blockquote>
+
+<h2 id="system-prompts"><strong>3. System Prompts & Personas</strong></h2>
+
+<p><strong>System prompt</strong> defines the model's role, behavior, constraints, and output format. Nó "sets the stage" trước khi user interaction.</p>
+
+<pre><code class="language-text">System Prompt:
+"You are a financial advisor AI for XYZ Bank.
+Rules:
+- Only answer questions about banking and investments
+- Never provide specific stock recommendations
+- Always include a disclaimer
+- Respond in a professional tone
+- If asked about non-financial topics, politely redirect"
+</code></pre>
+
+<h3 id="system-prompt-use"><strong>System Prompt Best Practices:</strong></h3>
+
+<table>
+<thead><tr><th>Practice</th><th>Why</th></tr></thead>
+<tbody>
+<tr><td>Define a <strong>clear role</strong></td><td>Constrains model behavior to a domain</td></tr>
+<tr><td>Set <strong>boundaries</strong></td><td>Prevents off-topic or harmful responses</td></tr>
+<tr><td>Specify <strong>output format</strong></td><td>Ensures consistent, parseable outputs</td></tr>
+<tr><td>Include <strong>examples</strong></td><td>Clarifies expected behavior</td></tr>
+<tr><td>Add <strong>guardrails</strong></td><td>Prevents misuse (PII, harmful content)</td></tr>
+</tbody>
+</table>
+
+<h2 id="advanced-techniques"><strong>4. Advanced Prompting Techniques</strong></h2>
+
+<h3 id="negative-prompting"><strong>4.1. Negative Prompting</strong></h3>
+
+<p>Chỉ rõ những gì model <strong>KHÔNG nên làm</strong>. Đặc biệt useful trong image generation.</p>
+
+<pre><code class="language-text">Text generation:
+"Summarize this article. Do NOT include opinions
+or personal commentary. Do NOT exceed 100 words."
+
+Image generation (Stable Diffusion):
+Prompt: "Professional headshot, studio lighting"
+Negative prompt: "blurry, cartoon, distorted, low quality"
+</code></pre>
+
+<h3 id="prompt-templates"><strong>4.2. Prompt Templates</strong></h3>
+
+<p>Reusable prompt structures với <strong>placeholders</strong> cho dynamic content:</p>
+
+<pre><code class="language-text">Template:
+"Given the following {document_type}:
+---
+{content}
+---
+Extract the following information:
+- {field_1}
+- {field_2}
+- {field_3}
+Respond in JSON format."
+</code></pre>
+
+<h3 id="prompt-chaining"><strong>4.3. Prompt Chaining</strong></h3>
+
+<p>Chia complex tasks thành <strong>multiple sequential prompts</strong>, output của prompt trước trở thành input của prompt sau.</p>
+
+<pre><code class="language-text">Step 1: "Extract key entities from this document: {doc}"
+         → Output: list of entities
+
+Step 2: "For each entity {entities}, find the sentiment
+         expressed about it in this text: {doc}"
+         → Output: entity-sentiment pairs
+
+Step 3: "Create a summary report of sentiment analysis
+         for these entities: {entity_sentiments}"
+         → Output: Final report
+</code></pre>
+
+<h2 id="comparison"><strong>5. Comparison Table for Exam</strong></h2>
+
+<table>
+<thead><tr><th>Technique</th><th>Examples Given?</th><th>Best For</th><th>Exam Keyword</th></tr></thead>
+<tbody>
+<tr><td><strong>Zero-shot</strong></td><td>None</td><td>Simple, well-known tasks</td><td>"no examples provided"</td></tr>
+<tr><td><strong>One-shot</strong></td><td>1 example</td><td>Setting format with minimal context</td><td>"single example"</td></tr>
+<tr><td><strong>Few-shot</strong></td><td>2-5 examples</td><td>Pattern following, classification</td><td>"examples provided", "demonstrations"</td></tr>
+<tr><td><strong>Chain-of-Thought</strong></td><td>With reasoning steps</td><td>Math, logic, complex reasoning</td><td>"step by step", "reasoning"</td></tr>
+<tr><td><strong>Negative prompting</strong></td><td>N/A</td><td>Avoiding unwanted outputs</td><td>"do not include", "avoid"</td></tr>
+<tr><td><strong>Prompt chaining</strong></td><td>N/A</td><td>Complex multi-step tasks</td><td>"break into steps", "sequential"</td></tr>
+</tbody>
+</table>
+
+<h2 id="inference-params"><strong>6. Inference Parameters Review</strong></h2>
+
+<p>Prompt engineering cũng bao gồm tuning inference parameters:</p>
+
+<table>
+<thead><tr><th>Parameter</th><th>Low Value</th><th>High Value</th></tr></thead>
+<tbody>
+<tr><td><strong>Temperature</strong></td><td>Deterministic, factual (0.0-0.3)</td><td>Creative, diverse (0.7-1.0)</td></tr>
+<tr><td><strong>Top-p</strong></td><td>Focused vocabulary (0.1-0.3)</td><td>Diverse vocabulary (0.9-1.0)</td></tr>
+<tr><td><strong>Top-k</strong></td><td>Limited choices (e.g., 10)</td><td>More choices (e.g., 250)</td></tr>
+<tr><td><strong>Max tokens</strong></td><td>Short responses</td><td>Long responses</td></tr>
+<tr><td><strong>Stop sequences</strong></td><td colspan="2">Define when to stop generating</td></tr>
+</tbody>
+</table>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "A customer support chatbot gives inconsistent answers" → Lower <strong>temperature</strong> (closer to 0). "A creative writing app produces boring text" → Raise <strong>temperature</strong> (closer to 1).</p>
+</blockquote>
+
+<h2 id="best-practices"><strong>7. Prompt Engineering Best Practices</strong></h2>
+
+<ol>
+<li><strong>Be specific</strong>: "Summarize in 3 bullet points" > "Summarize this"</li>
+<li><strong>Provide context</strong>: Include relevant background information</li>
+<li><strong>Define output format</strong>: JSON, markdown, table, bullet points</li>
+<li><strong>Use delimiters</strong>: Separate sections with --- or ``` to avoid prompt injection</li>
+<li><strong>Iterate</strong>: Test and refine prompts based on outputs</li>
+<li><strong>Avoid ambiguity</strong>: Don't assume model knows your intent</li>
+<li><strong>Use examples</strong>: When zero-shot doesn't work, add few-shot examples</li>
+</ol>
+
+<h2 id="practice-questions"><strong>8. Practice Questions</strong></h2>
+
+<p><strong>Q1:</strong> A developer is working on a classification task, but the model's zero-shot responses are inconsistent. Which prompting technique should the developer try NEXT?</p>
+<ul>
+<li>A) Reduce the temperature to 0</li>
+<li>B) Use few-shot prompting with example inputs and outputs ✓</li>
+<li>C) Fine-tune the model on custom data</li>
+<li>D) Switch to a different model provider</li>
+</ul>
+<p><em>Explanation: Few-shot prompting is the logical next step after zero-shot fails — providing examples helps the model understand the expected pattern. Fine-tuning is more expensive and complex. Temperature adjustment alone may not fix classification logic.</em></p>
+
+<p><strong>Q2:</strong> A customer wants their AI application to solve complex mathematical word problems more accurately. Which prompting technique would MOST improve the results?</p>
+<ul>
+<li>A) Zero-shot prompting</li>
+<li>B) Negative prompting</li>
+<li>C) Chain-of-Thought prompting ✓</li>
+<li>D) Prompt chaining</li>
+</ul>
+<p><em>Explanation: Chain-of-Thought prompting encourages the model to show its reasoning step by step, which significantly improves accuracy on mathematical and logical reasoning tasks.</em></p>
+
+<p><strong>Q3:</strong> Which of the following is a benefit of using a system prompt in a generative AI application?</p>
+<ul>
+<li>A) It eliminates the need for user input</li>
+<li>B) It reduces the model's inference cost</li>
+<li>C) It defines the model's role, behavior, and constraints ✓</li>
+<li>D) It replaces the need for fine-tuning</li>
+</ul>
+<p><em>Explanation: System prompts set the model's role, behavioral constraints, and output format — establishing consistent behavior across all user interactions without any model training.</em></p>

package/content/series/luyen-thi/luyen-thi-aws-ai-practitioner/chapters/03-domain-3-applications-foundation-models/lessons/06-bai-6-rag-vector-databases-knowledge-bases.md

@@ -0,0 +1,244 @@
+---
+id: 019c9619-lt01-d3-l06
+title: 'Bài 6: RAG, Vector Databases & Bedrock Knowledge Bases'
+slug: bai-6-rag-vector-databases-knowledge-bases
+description: >-
+  Retrieval-Augmented Generation (RAG) architecture.
+  Vector databases, embeddings, chunking strategies.
+  Amazon Bedrock Knowledge Bases. So sánh RAG vs Fine-tuning.
+duration_minutes: 60
+is_free: true
+video_url: null
+sort_order: 2
+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-bai6-rag-architecture.png" alt="RAG Architecture" style="max-width: 800px; width: 100%; border-radius: 12px;" />
+<p><em>RAG Architecture — Indexing Phase và Query Phase với Amazon Bedrock Knowledge Bases</em></p>
+</div>
+
+<h2 id="rag-overview"><strong>1. What is RAG?</strong></h2>
+
+<p><strong>Retrieval-Augmented Generation (RAG)</strong> là kỹ thuật kết hợp FM với <strong>external knowledge sources</strong> để trả lời chính xác hơn, giảm hallucination, và cập nhật thông tin mà model chưa biết.</p>
+
+<h3 id="why-rag"><strong>1.1. Why RAG?</strong></h3>
+
+<table>
+<thead><tr><th>Problem</th><th>RAG Solution</th></tr></thead>
+<tbody>
+<tr><td>Knowledge cutoff date</td><td>Retrieve latest documents</td></tr>
+<tr><td>Hallucination</td><td>Ground responses in real data</td></tr>
+<tr><td>No domain knowledge</td><td>Add company-specific documents</td></tr>
+<tr><td>Generic answers</td><td>Cite specific sources</td></tr>
+<tr><td>Privacy — can't send data to FM training</td><td>Keep data in your own vector DB</td></tr>
+</tbody>
+</table>
+
+<h3 id="rag-flow"><strong>1.2. RAG Architecture</strong></h3>
+
+<pre><code class="language-text">RAG Pipeline:
+
+┌─────────────────────────────────────────────────────────────┐
+│  INDEXING (Done once / periodically)                        │
+│                                                             │
+│  Documents → Chunking → Embedding Model → Vector Database   │
+│  (PDF, web,    (split     (Amazon Titan     (OpenSearch,    │
+│   S3, etc.)    text)       Embeddings)       Aurora pgvector)│
+└─────────────────────────────────────────────────────────────┘
+
+┌─────────────────────────────────────────────────────────────┐
+│  RETRIEVAL & GENERATION (Per query)                         │
+│                                                             │
+│  User Query → Embed Query → Search Vector DB → Top-K docs   │
+│                                                             │
+│  Augmented Prompt = System Prompt + Retrieved Docs + Query  │
+│                                                             │
+│  Augmented Prompt → Foundation Model → Answer with sources  │
+└─────────────────────────────────────────────────────────────┘
+</code></pre>
+
+<h2 id="chunking"><strong>2. Chunking Strategies</strong></h2>
+
+<p>Trước khi tạo embeddings, documents phải được <strong>chia nhỏ (chunked)</strong> thành các đoạn phù hợp.</p>
+
+<table>
+<thead><tr><th>Strategy</th><th>Description</th><th>Best For</th></tr></thead>
+<tbody>
+<tr><td><strong>Fixed-size</strong></td><td>Split every N characters/tokens</td><td>Simple, uniform documents</td></tr>
+<tr><td><strong>Sentence-based</strong></td><td>Split at sentence boundaries</td><td>Narrative text</td></tr>
+<tr><td><strong>Paragraph-based</strong></td><td>Split at paragraph breaks</td><td>Well-structured documents</td></tr>
+<tr><td><strong>Semantic</strong></td><td>Split based on topic changes</td><td>Complex documents</td></tr>
+<tr><td><strong>Hierarchical</strong></td><td>Parent-child chunk relationships</td><td>Long documents with sections</td></tr>
+</tbody>
+</table>
+
+<h3 id="chunk-size"><strong>Chunk Size Trade-offs:</strong></h3>
+
+<pre><code class="language-text">Small chunks (100-200 tokens):
+  ✓ More precise retrieval
+  ✗ May lose context
+  ✗ More chunks to search
+
+Large chunks (500-1000 tokens):
+  ✓ More context preserved
+  ✗ May include irrelevant info
+  ✗ Fewer chunks, less granular
+
+Overlap (e.g., 20% between chunks):
+  ✓ Prevents information loss at boundaries
+  ✗ Increases storage and compute
+</code></pre>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "How to improve RAG retrieval accuracy?" → Adjust <strong>chunk size</strong>, add <strong>overlap</strong>, use <strong>semantic chunking</strong>, improve <strong>embedding model</strong>.</p>
+</blockquote>
+
+<h2 id="embeddings"><strong>3. Embeddings for RAG</strong></h2>
+
+<h3 id="embedding-models"><strong>3.1. AWS Embedding Models</strong></h3>
+
+<table>
+<thead><tr><th>Model</th><th>Modality</th><th>Dimensions</th><th>Use Case</th></tr></thead>
+<tbody>
+<tr><td><strong>Amazon Titan Text Embeddings V2</strong></td><td>Text</td><td>256/512/1024</td><td>Semantic search, RAG</td></tr>
+<tr><td><strong>Amazon Titan Multimodal Embeddings</strong></td><td>Text + Image</td><td>256/384/1024</td><td>Cross-modal search</td></tr>
+<tr><td><strong>Cohere Embed</strong></td><td>Text</td><td>1024</td><td>Multilingual search</td></tr>
+</tbody>
+</table>
+
+<h3 id="vector-db"><strong>3.2. Vector Databases on AWS</strong></h3>
+
+<table>
+<thead><tr><th>Service</th><th>Type</th><th>Key Feature</th></tr></thead>
+<tbody>
+<tr><td><strong>Amazon OpenSearch Serverless</strong></td><td>Managed</td><td>Vector search collection type, serverless</td></tr>
+<tr><td><strong>Amazon Aurora PostgreSQL</strong></td><td>RDB + Vector</td><td>pgvector extension</td></tr>
+<tr><td><strong>Amazon Neptune</strong></td><td>Graph + Vector</td><td>Knowledge graphs with vector search</td></tr>
+<tr><td><strong>Amazon DocumentDB</strong></td><td>Document + Vector</td><td>MongoDB-compatible with vector search</td></tr>
+<tr><td><strong>Amazon MemoryDB</strong></td><td>In-memory + Vector</td><td>Redis-compatible, ultra-low latency</td></tr>
+<tr><td><strong>Pinecone (3rd party)</strong></td><td>Dedicated vector DB</td><td>Popular, integrates with Bedrock</td></tr>
+</tbody>
+</table>
+
+<h2 id="bedrock-kb"><strong>4. Amazon Bedrock Knowledge Bases</strong></h2>
+
+<p><strong>Bedrock Knowledge Bases</strong> là <strong>fully managed RAG solution</strong>. AWS handles chunking, embedding, indexing, retrieval — bạn chỉ cần point to data sources.</p>
+
+<h3 id="kb-architecture"><strong>4.1. How It Works</strong></h3>
+
+<pre><code class="language-text">Setup:
+┌───────────┐     ┌───────────────┐     ┌─────────────────┐
+│ S3 Bucket │────→│ Bedrock       │────→│ Vector Store     │
+│ (docs)    │     │ Knowledge Base│     │ (OpenSearch/     │
+│           │     │ (auto-chunk,  │     │  Aurora/Pinecone) │
+│           │     │  auto-embed)  │     │                  │
+└───────────┘     └───────────────┘     └─────────────────┘
+
+Query:
+┌───────────┐     ┌───────────────┐     ┌─────────────────┐
+│ User      │────→│ Knowledge Base│────→│ FM (Claude,      │
+│ "What is  │     │ retrieves     │     │  Titan, etc.)    │
+│  the..."  │     │ relevant docs │     │ generates answer │
+└───────────┘     └───────────────┘     └─────────────────┘
+</code></pre>
+
+<h3 id="kb-data-sources"><strong>4.2. Supported Data Sources</strong></h3>
+
+<ul>
+<li><strong>Amazon S3</strong>: PDF, TXT, MD, HTML, DOC, CSV</li>
+<li><strong>Web Crawler</strong>: Crawl websites automatically</li>
+<li><strong>Confluence</strong>: Atlassian Confluence pages</li>
+<li><strong>SharePoint</strong>: Microsoft SharePoint documents</li>
+<li><strong>Salesforce</strong>: Salesforce knowledge articles</li>
+</ul>
+
+<h3 id="kb-features"><strong>4.3. Key Features</strong></h3>
+
+<table>
+<thead><tr><th>Feature</th><th>Benefit</th></tr></thead>
+<tbody>
+<tr><td><strong>Managed chunking</strong></td><td>Auto-splits documents (fixed, semantic, hierarchical)</td></tr>
+<tr><td><strong>Auto-sync</strong></td><td>Periodically re-indexes when data changes</td></tr>
+<tr><td><strong>Source attribution</strong></td><td>Returns source documents with answers</td></tr>
+<tr><td><strong>Metadata filtering</strong></td><td>Filter chunks by custom metadata fields</td></tr>
+<tr><td><strong>Hybrid search</strong></td><td>Combines semantic + keyword search</td></tr>
+<tr><td><strong>Guardrails integration</strong></td><td>Apply safety filters to RAG responses</td></tr>
+</tbody>
+</table>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "A company wants to build a chatbot that answers questions from internal documents stored in S3, with minimal custom code" → <strong>Amazon Bedrock Knowledge Bases</strong>.</p>
+</blockquote>
+
+<h2 id="rag-vs-finetuning"><strong>5. RAG vs Fine-tuning</strong></h2>
+
+<table>
+<thead><tr><th>Factor</th><th>RAG</th><th>Fine-tuning</th></tr></thead>
+<tbody>
+<tr><td><strong>Purpose</strong></td><td>Access external/current data</td><td>Teach new skills/domain patterns</td></tr>
+<tr><td><strong>Data freshness</strong></td><td>Always up-to-date</td><td>Fixed at training time</td></tr>
+<tr><td><strong>Training required?</strong></td><td>No model training</td><td>Yes, needs labeled data + compute</td></tr>
+<tr><td><strong>Cost</strong></td><td>Vector DB + retrieval costs</td><td>Training compute + storage</td></tr>
+<tr><td><strong>Hallucination</strong></td><td>Reduced (grounded in data)</td><td>May still hallucinate</td></tr>
+<tr><td><strong>Latency</strong></td><td>Slightly higher (retrieval step)</td><td>Same as base model</td></tr>
+<tr><td><strong>Best for</strong></td><td>Q&A, search, knowledge bases</td><td>Style, tone, domain-specific patterns</td></tr>
+<tr><td><strong>Data privacy</strong></td><td>Data stays in your vector DB</td><td>Data used in training process</td></tr>
+</tbody>
+</table>
+
+<h3 id="when-to-use"><strong>Decision Matrix:</strong></h3>
+
+<pre><code class="language-text">"Need to answer from company docs?"       → RAG
+"Need real-time/latest information?"       → RAG
+"Need to change model's writing style?"    → Fine-tuning
+"Need model to follow specific format?"    → Try prompting first → then fine-tuning
+"Need domain-specific terminology?"        → RAG (if in docs) or Fine-tuning (if patterns)
+"Minimum effort/cost?"                     → RAG > Prompt Engineering > Fine-tuning
+</code></pre>
+
+<h2 id="rag-evaluation"><strong>6. Evaluating RAG Quality</strong></h2>
+
+<table>
+<thead><tr><th>Metric</th><th>What it measures</th></tr></thead>
+<tbody>
+<tr><td><strong>Faithfulness</strong></td><td>Is the answer grounded in retrieved docs? (no hallucination)</td></tr>
+<tr><td><strong>Relevance</strong></td><td>Are retrieved documents relevant to the query?</td></tr>
+<tr><td><strong>Answer correctness</strong></td><td>Is the final answer factually correct?</td></tr>
+<tr><td><strong>Context precision</strong></td><td>What % of retrieved chunks are actually relevant?</td></tr>
+<tr><td><strong>Context recall</strong></td><td>Did we retrieve all relevant chunks?</td></tr>
+</tbody>
+</table>
+
+<h2 id="practice-questions"><strong>7. Practice Questions</strong></h2>
+
+<p><strong>Q1:</strong> A healthcare company wants an AI assistant that answers questions from their latest medical research papers stored in Amazon S3. The information changes weekly. Which approach is MOST suitable?</p>
+<ul>
+<li>A) Fine-tune a foundation model on the papers</li>
+<li>B) Use RAG with Amazon Bedrock Knowledge Bases ✓</li>
+<li>C) Use zero-shot prompting with a large context window</li>
+<li>D) Pre-train a custom model on medical data</li>
+</ul>
+<p><em>Explanation: RAG with Bedrock Knowledge Bases is ideal — it automatically indexes S3 documents, retrieves relevant information per query, and keeps responses current without retraining. Weekly updates are handled by auto-sync.</em></p>
+
+<p><strong>Q2:</strong> What is the PRIMARY purpose of chunking documents in a RAG pipeline?</p>
+<ul>
+<li>A) To reduce storage costs</li>
+<li>B) To split documents into manageable pieces for embedding and retrieval ✓</li>
+<li>C) To encrypt sensitive data</li>
+<li>D) To convert documents to a different file format</li>
+</ul>
+<p><em>Explanation: Chunking splits large documents into smaller, semantically meaningful pieces that can be individually embedded and retrieved. This enables precise retrieval of relevant information rather than processing entire documents.</em></p>
+
+<p><strong>Q3:</strong> A company built a RAG application, but it sometimes returns answers not supported by the retrieved documents. Which metric should they focus on improving?</p>
+<ul>
+<li>A) Context recall</li>
+<li>B) Answer length</li>
+<li>C) Faithfulness ✓</li>
+<li>D) Response latency</li>
+</ul>
+<p><em>Explanation: Faithfulness measures whether the generated answer is grounded in the retrieved documents. Low faithfulness means the model is generating information beyond what the retrieved context supports (hallucination in RAG context).</em></p>