RAG vs Fine-Tuning

What problem do you actually have?

Pick the closest one.

Use RAG

The facts are missing, and fine-tuning does not add a searchable knowledge base. Retrieve the right passages at query time and pass them in.

Learn RAG

The decision on one page

Which fix do you need?Start at the top with the symptom, not the technology.

flowchart TD
  A{"What is wrong with the output?"} -->|"Missing or stale facts"| B["RAG"]
  A -->|"Format or instructions"| C["Prompting and few-shot examples"]
  A -->|"Behavior prompts cannot hold"| D{"Do you have labeled examples and an eval?"}
  D -->|No| E["Build the eval set first"]
  D -->|Yes| F["Fine-tuning"]
  C --> G{"Fixed?"}
  G -->|No| D
  B --> H{"Also a style problem?"}
  H -->|Yes| D

When each one does its workRAG acts on every request. Fine-tuning happens before release.

flowchart LR
  subgraph R["RAG: at every request"]
    R1["Question"] --> R2["Search your index"] --> R3["Add passages"] --> R4["Base model"]
  end
  subgraph T["Fine-tuning: once, before release"]
    T1["Examples"] --> T2["Training job"] --> T3["Tuned model"]
  end

Side by side

	RAG	Fine-tuning	Prompting
What it changes	What the model can read for this request	How the model tends to respond	What you ask for
Best for	Facts, documents, anything that changes	Style, format habits, classification boundaries	Most problems, first
Updating it	Re-index the content	Prepare data and train again	Edit text and re-run your evals
Added to each request	Retrieval step and extra input tokens	Usually none, beyond the tuned model	Longer prompt if you add examples
Typical failure	Wrong or missing passages	Overfits, or goes stale with its base model	Instructions followed unevenly
You need	An index and retrieval evals	Labeled examples and evals	A test set

Where fine-tuning stands now

Availability changed a lot this year, so check your provider before you plan around it.

OpenAI

Winding down. The platform is closed to new users, and existing customers can create new jobs until January 6, 2027. Existing tuned models keep working until their base models are deprecated.

Amazon Bedrock

Offered. AWS lists supervised fine-tuning, reinforcement fine-tuning and distillation.

Microsoft Foundry

Offered, including supervised fine-tuning through Managed Training.

Choose RAG for knowledge

RAG retrieves documents at query time and grounds the model in sources. It is best for current data, enterprise documents, citations, policies, product catalogs, and fast-changing facts.

Choose fine-tuning for behavior

Fine-tuning changes model behavior through examples. It is best for style, classification boundaries, structured task habits, domain-specific formats, and repeated patterns that prompts cannot reliably enforce.