Prompting vs RAG vs Fine-Tuning vs Tools

The right intervention follows the type of gap: instructions, knowledge, behavior, or action.

Use prompting for task specification

Prompts define goals, constraints, examples, and output contracts. They are the lowest-complexity way to change behavior at request time, but they cannot grant access to unknown or current private information.

Use RAG for governed knowledge

Retrieval supplies updateable, attributable evidence at inference time. It introduces ingestion, indexing, access-control, retrieval-quality, and citation responsibilities. It is not a substitute for stable behavior training.

Use fine-tuning for demonstrated behavior gaps

Fine-tuning can improve stable style, task patterns, or efficiency after a representative dataset and evaluation demonstrate the need. Facts that change frequently are poor training targets.

Use tools for truth and action

Databases, calculators, APIs, and business services should provide live state, deterministic computation, and side effects. The model proposes; application code validates, authorizes, executes, and records.

Combine with an explicit hypothesis

A support system may use prompting for tone, RAG for policy, tools for account state, and fine-tuning for persistent classification errors. Add each mechanism only when it addresses a measured failure and survives its own risk review.

Decision checklist

1. Name the user outcome and unacceptable failures.
2. Identify the layer where the observed problem originates.
3. Choose the smallest mechanism that directly addresses that problem.
4. Define representative evaluation cases and operational budgets.
5. Preserve source, model, prompt, data, and release versions.
6. Require explicit approval before changing public behavior.