The article reframes the common RAG versus fine-tuning debate by showing they operate at different layers of an AI stack. RAG retrieves external documents at inference time, converts them to embeddings using models like text-embedding-3-small, stores them in a vector database, and injects relevant chunks into the prompt. The base model — such as gpt-4o-mini — remains unchanged. This approach excels at answering questions about private or rapidly changing data, providing citable sources, and avoiding retraining costs.

Fine-tuning instead updates model weights through continued training on task-specific input-output pairs formatted as JSONL. The OpenAI API example shows a minimum of 50-100 examples needed to produce a distinct model identifier (ft:gpt-4o-mini-2024-07-18:org:suffix:id). This modifies the model's behavior, tone, reasoning style, and output formatting — things RAG cannot address.

The practical distinction: use RAG when the challenge is knowledge access — new documents, proprietary data, or freshness requirements. Use fine-tuning when the model consistently fails at a specific task format, style, or reasoning pattern that retrieval cannot fix. Many production systems combine both: a fine-tuned model for consistent behavior, augmented by RAG for domain knowledge.