Making it "Yours"

A standard foundation model (FM) knows a lot about the world, but it knows nothing about your company.

• It doesn't know your specific employee handbook.
• It doesn't know your inventory from yesterday.
• It doesn't know your secret project codenames.

To fix this, we use two primary integration patterns: RAG and Fine-Tuning. On the AWS Certified AI Practitioner exam, you must be able to choose between these two based on the "Dynamic nature" and "Privacy" of the data.

---

1. RAG: Retrieval-Augmented Generation (The "Open-Book" Test)

Definition: RAG is a technique where you give the AI a "Search Tool" so it can look up fresh information before it answers your question.

Think of it like an Open-Book Exam.

• The AI doesn't need to memorize your manual.
• When a user asks a question, the system searches your document library (The "Retrieval" bit), finds the relevant 2 paragraphs, and sends them to the AI along with the prompt.
• The AI "Augments" its knowledge with those paragraphs and generates an answer.

Why use RAG?

• Freshness: If your data changes every 5 minutes (like stock prices), RAG is the only way.
• Accuracy: It virtually eliminates hallucinations because you can force the AI to cite its sources.
• Cheap: You don't have to "Train" anything.

AWS Service: Knowledge Bases for Amazon Bedrock

AWS has a feature that automates the whole RAG pipeline. You just point Bedrock to an S3 Bucket, and it handles the rest.

---

2. Fine-Tuning (The "Specialist" Education)

Definition: Fine-tuning is the process of taking a "Junior" foundation model and training it on a small, high-quality dataset of your own to make it a "Senior" expert.

Think of it like Medical School.

• You take someone who already knows English (The FM) and you have them read 10,000 medical textbooks until they learn the "Voice" and "Acronyms" of a doctor.
• The knowledge is now "Baked Into" the model itself.

Why use Fine-Tuning?

• Style and Tone: If you want your AI to talk exactly like your brand's unique "voice."
• Niche Domains: For specialized scientific or legal data that requires deep, internalized understanding.
• Static Knowledge: For rules that rarely change but are very complex.

---

3. Comparison for the Exam

Feature	RAG (Retrieval)	Fine-Tuning (Training)
Logic	Look it up in a book.	Memorize the book.
Data Update	Near-Instant.	Slow (Needs re-training).
Hallucination	Low (Cites sources).	Medium (Can still hallucinate).
Effort	Low to Medium.	High (Needs clean data).

graph TD
    subgraph Patterns
    A[Human Prompt] --> B{Strategy}
    
    B -->|RAG| C[Search S3/Database]
    C --> D[Retrieve Relevant Text]
    D --> E[Combine Prompt + Text]
    E --> F[Generate Grounded Answer]
    
    B -->|Fine-Tuning| G[Build Specialized FM]
    G --> H[Generate Expert-Tone Answer]
    end

---

4. Summary: The 90/10 Rule

In 2026, 90% of business problems should be solved with RAG. Fine-tuning is expensive, difficult, and the model "forgets" things over time. Only use fine-tuning if you need a specific "Tone" or if RAG literally cannot provide enough context for the model to understand the domain.

---

Exercise: Choose the Integration

A real-estate company wants an AI chatbot that can answer questions about "Current Listings" which change every hour as houses are bought and sold. They also want the AI to cite the URL of the house it is talking about. Which pattern should they use?

• A. Fine-Tuning.
• B. RAG (Retrieval-Augmented Generation).
• C. Prompt Engineering only.
• D. Traditional SQL Querying.

The Answer is B! Because the data is Dynamic (changes every hour) and requires Citations (URLs), RAG is the perfect solution.

---

Knowledge Check

What's Next?

We know the patterns. But why Bedrock? Why not just run these models on your own servers? In our final lesson of the module, we look at the Benefits of fully managed foundation models.