Prompt, RAG, Fine-Tune, or Tool?

Outcome. Given a measured failure, you can choose prompting, retrieval, fine-tuning, deterministic code, or a tool and state the eval that would prove the choice worked.

Different interventions solve different problems. Prompting changes the instructions and examples supplied at inference, making it the fastest first lever for task definition, boundaries, and format. RAG supplies changing, private, or sourceable knowledge. A tool performs a calculation, lookup, or action that should come from an external system. Fine-tuning updates model parameters from examples and can improve recurring behavior, classification, style, domain task performance, or smaller-model economics. Deterministic code is better when a rule can be computed exactly.

Fine-tuning is not a dependable database for facts that change weekly; maintaining retrieval or calling the source system is usually easier to update and audit. RAG will not fix a weak decision policy. A calculator tool is better than hoping a model memorizes arithmetic. A larger prompt is not always better than a concise contract plus tested examples.

Start from an eval and error analysis. Establish a baseline, categorize failures, try the least expensive reversible intervention, and retest. Combine levers when the failure spans layers—for example, retrieval for current policy, a prompt for evidence use, and code for refund limits.

Mental model. Prompt for behavior, retrieve for changing evidence, fine-tune recurring learned behavior, use tools for external truth/action, and use code for exact rules.

Evidence trail — reviewed 23 July 2026. InstructGPT: https://arxiv.org/abs/2203.02155. LoRA: https://arxiv.org/abs/2106.09685. QLoRA: https://arxiv.org/abs/2305.14314. RAG: https://arxiv.org/abs/2005.11401. Always begin with a task-specific evaluation.