Pergunta comum em entrevista de AI / LLM Engineering. A resposta abaixo está em inglês, do jeito que dá para falar numa entrevista: leia, feche e tente explicar com as suas palavras.
Prompt engineering changes instructions, examples, and context without training model weights. It is often a useful first experiment, but its token and evaluation costs still matter.
RAG supplies retrieved knowledge at query time. Updating sources and indexes can keep facts current without retraining and provide evidence for citations.
Fine-tuning changes weights to improve behavior or task performance and can also teach facts. Maintaining rapidly changing facts and traceable sources is usually easier through retrieval.
Full fine-tuning updates all model parameters. LoRA freezes the base weights and trains low-rank updates, reducing trainable parameters. Total cost and quality still depend on model, data, and configuration.
For changing company documents that require citations, evaluate retrieval first. Fine-tuning may complement it when behavior also needs improvement.
Source: Original LoRA paper.
Ler é o primeiro passo. No app, essa pergunta vira flashcard com repetição espaçada, e o módulo tem quiz e simulado para testar o que ficou.
Treinar essa pergunta no app