Prompt Engineering vs Finetuning

When to use one vs the other?

Prompt Engineering

• Fast to change (if using a Prompt Store)
• Less work to upgrade when the underlying model improves
• Can get good performance if you experiment
• Augments the model with factual content to mitigate “hallucinations”

Finetuning

• Have access to a special dataset (e.g. user or customer data) and want better performance
• Improved latency (10x or more)
• Improved cost
• Trying to achieve a particular tone
• Local/private model

In general, start with prompt engineering and push that as far as you can. Only then, look to finetuning for further optimisation.