Three Levers and What Each One Changes

Instructions, knowledge, behaviour

Prompting changes the input (instructions, examples, format); the model is untouched, so changes are instant and reversible. Retrieval (RAG) also changes the input, adding knowledge the model lacks at question time. Fine-tuning changes the model itself: continued training on your examples shifts its weights and so its default behaviour (style, format, narrow skills). Missing or changing facts point to retrieval, unclear instructions to prompting, and consistent behaviour a prompt cannot reliably produce to fine-tuning. They combine well: a tuned model can still use retrieval.

Change the input, add knowledge, or change the model

Three levers change an LLM system in different ways and at very different costs.

Which lever for which problem

A first-pass guide; always confirm with an evaluation.

Symptom                                                Likely lever
answers lack our private / recent facts                 RAG (retrieval)
model ignores or misreads instructions                  better prompt, examples, structure
format/tone drifts despite clear instructions            few-shot first; then fine-tune
prompt is huge (many examples) and costs too much        fine-tune a smaller model
need a narrow skill a general model does poorly          fine-tune (with enough good data)
model does not know the topic at all                     stronger model and/or retrieval

Name the gap first

Write down whether the failure is facts, instructions or behaviour before picking a lever; most wasted fine-tuning starts with a skipped diagnosis.

Quick check: Which lever adds fresh, changing facts most reliably?

• A longer style prompt
• Fine-tuning
• Retrieval (RAG)
• Raising temperature