Learning outcomes

  • Diagnose the actual gap
  • Choose an adaptation strategy
  • Combine strategies intentionally

Mental model

Prompting changes instructions, RAG supplies external knowledge, fine-tuning changes learned behavior, and tools provide actions or authoritative computation.

Observed failure
Knowledge behavior or action gap
Select intervention
Build evaluation
Combine only when justified

Theory

Diagnose the failure first. Missing current or private knowledge points toward retrieval. Inconsistent style or stable task behavior may justify fine-tuning after prompts and evaluations. Calculations, live state, and side effects belong in tools. These strategies combine: a fine-tuned model can call tools over retrieved evidence.

Alternatives and trade-offs

Long context can replace retrieval for small corpora, deterministic code can replace model reasoning for known rules, and a better base model may outperform adaptation complexity.

Failure modes and misconceptions

Do not fine-tune facts that change often, use RAG to enforce tone, call a model for deterministic arithmetic, or choose from fashion rather than measured constraints.

Knowledge check

Reflect before revealing the guide

Which intervention best addresses a model that cannot access today's account balance?

Decision scenario

A support assistant uses prompting for tone, RAG for policy, a billing tool for live balances, and fine-tuning only if labeled evaluations show a persistent behavior gap.

Relationships

Primary sources