The instinct to fine-tune a model is often a symptom of an under-engineered prompt or an empty retrieval layer. Each tool in the adaptation ladder costs an order of magnitude more than the previous one.
The escalation ladder
Start with prompt engineering: instructions, examples and output schemas solve the majority of behavior issues. Escalate to retrieval augmentation when the missing piece is knowledge rather than behavior. Fine-tune only when you need consistent style, strict format reliability or domain behavior that prompting cannot hold.
Signals you actually need fine-tuning
Thousands of verified input-output pairs exist. Latency budgets forbid long prompts. The task is narrow and repeated at scale. Output format failures persist despite disciplined prompting.
Fine-tuning is a data project wearing an AI costume. The dataset, not the GPU, is the bottleneck.
The economics
Prompt iteration takes hours and zero infrastructure. Retrieval takes weeks and a document pipeline. Fine-tuning takes months of dataset curation plus evaluation loops and retraining whenever the product shifts.
Our rule at ByteHexa: exhaust the cheap layers, measure, and only then climb. Most projects never need the expensive ones, and the ones that do arrive with the data to justify it.
