Project 42
Start learning
Module 1 of 5 · 40 min

The AI Customization Decision Matrix

Evaluate when to use Prompt Engineering, Few-Shot In-Context Learning, RAG, PEFT/LoRA, or Full Fine-Tuning.

Core concept

By the end

You will be able to

  • Apply the 4-tier customization framework: Prompting vs RAG vs LoRA vs Continual Pre-training.
  • Calculate cost, maintenance overhead, and latency impacts for each customization layer.
  • Identify task requirements that necessitate weight modification versus context injection.
01

Knowledge vs Form & Style

Fine-tuning is for modifying style, tone, format, and domain jargon. RAG is for injecting dynamic facts, access-controlled knowledge, and up-to-the-minute data.

Using fine-tuning to teach a model rapidly changing enterprise facts leads to catastrophic forgetting, high retraining costs, and hallucinated outdated answers.

Customization Selection Framework
text
Need Dynamic Facts? -> Use RAG
Need Strict Output Format / Domain Tone? -> Use LoRA Fine-Tuning
Need Custom Vocabulary / Specialized Language? -> Use Continual Pre-Training
Need Task Guidance? -> Use Prompt Engineering & Few-Shot In-Context

Practice activity

Architect Customization Strategy for Three Enterprise Scenarios

  1. Analyze 3 enterprise requirements: Real-time stock ticker analysis, medical diagnostic report formatting, and proprietary code translation.
  2. Map each scenario to the optimal customization tier with full architectural justification.

What to produce

  • Architecture decision record (ADR) detailing customization choices and cost estimates.

Reflect before continuing

Why is fine-tuning poorly suited for real-time inventory lookups?

Evidence

Sources and verification

Knowledge check

Make it stick.

Pass at 80%

Choose the strongest answer for each question. Your attempts become part of your account transcript.

01When is LoRA fine-tuning significantly preferred over RAG?