Skip to main content
📊

Fine-tuning Decision Tree

When to fine-tune vs prompt vs RAG

1 page·410 KB·CC-BY 4.0
fine-tuningragpromptingdecision-makingoptimization

View Resource

No email required. No signup. View online or print as PDF.

View Full Resource →

What's included

  • Visual decision tree with yes/no questions
  • Guides you to: Prompting, RAG, or Fine-tuning
  • Includes cost and complexity comparison
  • Real-world examples for each path
  • Printable flowchart format
  • Perfect for technical planning and architecture decisions

Why you need this

You need better AI outputs. But should you:

  • Write better prompts?
  • Build a RAG system?
  • Fine-tune a model?

This decision tree gives you the answer in 60 seconds.

Perfect for:

  • Developers deciding how to customize AI models
  • Product managers scoping AI features
  • CTOs evaluating technical approaches
  • Teams trying to avoid over-engineering (or under-engineering)

What's inside

The Decision Tree

Answer 5 simple questions:

  1. Do you need the AI to know specific facts or data?

  2. Is your data public or commonly known?

  3. Do you need a specific writing style, tone, or format?

  4. Is cost a major concern? (Fine-tuning = $$$)

    • Yes → Try prompting or RAG first
    • No → Fine-tuning is an option
  5. How often does your data change?

    • Constantly → Use RAG (easy to update)
    • Rarely → Fine-tuning might work

The Three Paths

Path 1: Better Prompting

  • When: You need general knowledge or creative tasks
  • Cost: Free to cheap
  • Effort: Low (minutes to hours)
  • Examples: Drafting emails, brainstorming, summarizing

Path 2: RAG (Retrieval-Augmented Generation)

  • When: You need AI to access your own data (docs, databases, FAQs)
  • Cost: Moderate (vector DB + API calls)
  • Effort: Medium (days to weeks)
  • Examples: Customer support bots, document Q&A, internal knowledge bases

Path 3: Fine-tuning

  • When: You need a specific style, tone, or highly specialized behavior
  • Cost: High (training + infrastructure)
  • Effort: High (weeks to months)
  • Examples: Medical diagnosis, legal document generation, brand-specific writing

Cost & Complexity Comparison Table

Approach Cost Effort Update Speed Use Case
Prompting $ Hours Instant General tasks
RAG $$ Days Fast (just update docs) Custom data access
Fine-tuning $$$ Weeks Slow (requires retraining) Specialized behavior

How to use it

  • Start here before committing to a technical approach
  • Save time and money by avoiding unnecessary fine-tuning
  • Share with stakeholders to explain technical trade-offs
  • Revisit quarterly — as your needs evolve, your approach might change

Real-world examples

Example 1: E-commerce Chatbot

  • Need: Answer product questions from your catalog
  • Decision: RAG (constantly updating inventory)
  • Why not fine-tuning? Too slow to retrain every time products change

Example 2: Legal Document Generator

  • Need: Generate contracts in your firm's exact style and format
  • Decision: Fine-tuning (specific style + compliance requirements)
  • Why not RAG? Style and tone are more important than facts

Example 3: Marketing Copy

  • Need: Draft social media posts and blog intros
  • Decision: Better prompting (general task, doesn't require custom data)
  • Why not fine-tuning? Overkill for a simple creative task

Want to go deeper?

This decision tree is your starting point. For implementation guides and advanced techniques:

License & Attribution

This resource is licensed under Creative Commons Attribution 4.0 (CC-BY). You're free to:

  • Share with your engineering team
  • Print for architecture planning sessions
  • Include in technical proposals

Just include this attribution:

"Fine-tuning Decision Tree" by Field Guide to AI (fieldguidetoai.com) is licensed under CC BY 4.0

Download now

Click below for instant access. No email required.

Related Guides

Key Terms

Ready to view?

Access your free Fine-tuning Decision Tree now. No forms, no wait—view online or print as PDF.

View Full Resource →

Licensed under CC-BY 4.0 · Free to share and adapt with attribution