方寸 Portal

AI Engineer Portal

Your personal operating system for career transition.

Private mode

Knowledge Note

A decision guide for prompting, RAG, and fine-tuning

Pick the right technique based on knowledge freshness, control needs, and operational cost.

Category

concept

Tags

prompting · rag · fine-tuning

Sources

1 linked references

Prompting vs RAG vs fine-tuning

Use prompting when behavior is the main variable, RAG when knowledge changes, and fine-tuning when output behavior itself must become more reliable.

This decision is easier when you stop asking "Which technique is most advanced?" and start asking:

  • what is changing?
  • what must stay controlled?
  • what will be expensive to maintain?

Prompting

Prompting is the best first move when:

  • the model already knows enough
  • the main need is better instructions or output structure
  • the workflow is still evolving quickly

Use prompting for:

  • formatting
  • role framing
  • output constraints
  • lightweight behavioral steering

RAG

RAG is the better move when:

  • the needed knowledge is external or changing
  • the answer must cite real source material
  • you need to ground outputs in specific documents

Use RAG when freshness and provenance matter more than memorized behavior.

Fine-tuning

Fine-tuning is worth considering when:

  • the same behavior needs to become consistently stronger
  • prompting alone is too brittle
  • retrieval will not solve the problem because the issue is output behavior rather than missing knowledge

It is not the first answer to every weak AI feature.

Decision table

Choose prompting when:

  • instructions are the main issue
  • you want maximum iteration speed

Choose RAG when:

  • knowledge changes often
  • citation and grounding matter

Choose fine-tuning when:

  • output behavior itself needs to change
  • you have enough examples to justify the maintenance cost

Common mistakes

  • using RAG to solve a pure formatting problem
  • trying to fine-tune around missing external knowledge
  • endlessly prompt-tuning when the real issue is retrieval quality

Practical takeaway

The best choice is the one that isolates the real failure mode with the least long-term complexity.