Fine-Tuning, Retrieval and Memory Solve Three Different Problems

Almost every team that ships an LLM feature runs into the same moment. The answers are wrong often enough to matter, someone says the model needs to be trained on our data, and a quarter of the budget