AWS Certified AI Practitioner AIF-C01 Practice Question

Retrieval-Augmented Generation (RAG) keeps the foundation model frozen and augments each prompt with relevant passages retrieved from an external knowledge base or vector index. Because no model parameters are updated, there is no need to allocate GPU time for training, making the upfront cost very low.

Pre-training a model from scratch is the most expensive option because it requires large datasets, lengthy compute time, and extensive engineering effort. Fully fine-tuning the base model is also costly due to the need for GPU resources and storage of a new model copy. Parameter-efficient fine-tuning (for example, using adapters or LoRA) reduces the amount of trainable parameters but still involves a training phase, so it incurs more upfront cost than a RAG workflow.