도메인 지식을 위한 미세 조정과 검색 비교

To make an LLM expert in your domain you can:

• Fine-tune: adjust the model weights on your data
• Retrieve: fetch relevant documents at query time (RAG)

Often the best answer is a combination of both.

Fine-tuning bakes patterns into the weights. It is excellent for teaching style, format, and behavior — for example always replying in a strict JSON shape or a brand voice.

Retrieval injects facts at runtime without touching weights. It shines when knowledge is large, frequently updated, or must be cited to a source.

A fine-tuned model freezes knowledge at training time. If your facts change daily (prices, policies, inventory), retrieval wins because you just update the document store.

Fine-tuning has upfront training cost and a new model to maintain. Retrieval has ongoing per-query embedding and storage cost but no retraining. Match the model to your update cadence.

Fine-tuning data is usually prompt/response pairs in JSONL. The model learns to imitate the desired responses.

RAG embeds the query, finds nearby chunks, and stuffs them into the prompt as context.

Fine-tuning on facts can increase hallucination — the model confidently states learned facts even when wrong. Retrieval lets you cite and verify, reducing fabrication.

A powerful pattern: fine-tune for behavior and format, use retrieval for current facts. The model knows how to act; the documents tell it what is true today.

Fine-tuning needs hundreds to thousands of clean, consistent examples. Retrieval needs only well-chunked documents and embeddings — far less labeling effort to start.

Choose retrieval if knowledge is large, changing, or must be cited. Choose fine-tuning if you need consistent style/format or lower per-call latency on a fixed behavior. Combine them when you need both.

Test your understanding.

You compared fine-tuning (best for style, format, behavior) with retrieval (best for large, fresh, citable facts). Fine-tuning can worsen factual hallucination; combining a fine-tuned behavior model with retrieval for facts is often the strongest domain customization.