Fine-tuning vs prompt engineering vs RAG — decision guide?

Go in order of cost and complexity, and only move up when you've measured that the previous step isn't enough:

1. Prompt engineering + examples (hours): a clear system prompt describing the brand voice, plus 5–10 excellent example descriptions (few-shot) in each language. This gets most teams 80–90% of the way. 2. RAG / dynamic examples (days): retrieve the most similar existing descriptions (same category) and include them as examples. That helps consistency and terminology, for instance your Kreyòl glossary. 3. Fine-tuning (weeks, plus data work): only when prompting plateaus, you have hundreds or thousands of high-quality examples, and you need consistency at high volume or lower cost with a smaller model.

For Haitian Creole specifically: test several models early. Quality in lower-resource languages varies a lot. Have native speakers review a sample, and keep a glossary of preferred terms (spelling standardization matters).

Whatever you choose, build a small evaluation set judged by humans so you can compare approaches objectively.