Retrieval or fine-tuning: choose based on what must change
Use retrieval when the system needs access to changing facts, private documents, or source citations. Use fine-tuning when the goal is to make behavior, style, formatting, or task execution more consistent.
The distinction matters because the two methods solve different problems. A tuned model does not become a reliable database of changing company policy, while a retrieval layer does not automatically teach a model a new response style or decision boundary.
Many production systems need both: retrieval for current knowledge, a tuned or carefully prompted model for behavior, and evaluations that test the combined pipeline. The wrong question is “Which technique is better?” The useful question is “What exactly must improve, and how often will it change?”
- Category
- Research
- Platform
- Web