Fine-Tuning vs Prompting vs RAG: Choosing the Right Tool

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ai-foundations-fine-tuning-final1-creators

1. What percentage of accuracy improvements can typically be achieved through better prompting alone?

   1. 10%
   2. 20%
   3. 50%
   4. 80%

2. A developer notices their AI frequently gives outdated information about company policies that change weekly. Which approach would best address this issue?

   1. Implement RAG with up-to-date policy documents
   2. Fine-tune the model on the latest policies
   3. Train a new model from scratch
   4. Add more examples to the prompt

3. A team needs their AI to always respond in a specific JSON format for their software to parse correctly. Which method is most appropriate?

   1. Training a brand new model optimized for JSON
   2. Using RAG to retrieve format templates
   3. Fine-tuning for consistent output format
   4. Writing longer prompts describing the format

4. A healthcare organization wants to add private patient data to their AI without risking that data being exposed through the model. Which approach is safer?

   1. Training from scratch is the safest option
   2. RAG keeps data separate from model weights
   3. Both approaches are equally safe
   4. Fine-tuning embeds data safely

5. What happens to a fine-tuned model when its base model is deprecated by the AI provider?

   1. The fine-tune continues working indefinitely
   2. The fine-tune becomes more accurate
   3. The fine-tune automatically migrates to the new base
   4. The fine-tuned model also becomes unusable

6. A startup wants to make their large AI model behave like a smaller, cheaper one while keeping similar capabilities. What technique accomplishes this?

   1. Using RAG with the large model
   2. Prompt engineering the large model
   3. Fine-tuning the large model repeatedly
   4. Distilling the large model's behavior into a smaller model

7. A student has a dataset full of incorrect information. Can fine-tuning fix this problem?

   1. No — you cannot fine-tune your way out of fundamentally bad data
   2. Yes — fine-tuning will correct the errors automatically
   3. Yes — but only with more training data
   4. No — but RAG can fix bad data

8. A company needs to update their AI with new product information daily. Which approach requires the least ongoing maintenance?

   1. Fine-tuning — just retrain the model
   2. Training from scratch — most reliable
   3. RAG — just update the knowledge base
   4. Prompt engineering — just update the prompt text

9. A developer wants to know when prompting alone is sufficient for their project. What signal indicates prompting is enough?

   1. The model already answers correctly with better prompts
   2. The model needs to access private documents
   3. The model outputs vary too much in style
   4. The model needs to learn a new language

10. A legal firm wants their AI to always use formal, precise language matching their court filings. What is the most appropriate solution?

    1. Add more legal terms to the system message
    2. Fine-tune for consistent formal style
    3. Use RAG to retrieve legal templates
    4. Write prompts emphasizing formality

11. Why is fine-tuning considered a last resort in the improvement hierarchy?

    1. It requires no technical expertise
    2. It is expensive, inflexible, and creates a fixed artifact
    3. It always produces better results than prompting
    4. It is too easy to implement correctly

12. A retail company wants to add comprehensive product catalog information to their AI assistant. Which method efficiently adds this domain knowledge?

    1. Prompt engineering lists every product
    2. Training a custom model from scratch
    3. Fine-tuning embeds all product details
    4. RAG connects to product catalog documents

13. What evidence would justify moving from prompting to RAG?

    1. The model outputs vary slightly between responses
    2. The model occasionally misses a minor detail
    3. The model takes too long to generate responses
    4. The model needs access to information beyond what fits in a prompt

14. A team is deciding between RAG and fine-tuning for their knowledge base application. What factor favors RAG?

    1. They need a specific writing style
    2. They want to distill to a smaller model
    3. They need consistent JSON output
    4. Their knowledge base updates frequently

15. What happens to the knowledge in a fine-tuned model when new information becomes available?

    1. The model selectively forgets old information
    2. The model retains its original training and cannot learn new information without retraining
    3. The model becomes more accurate over time automatically
    4. The model automatically incorporates new knowledge