RAG vs Fine-Tuning on AWS — When to Use Each
A frequent AWS AI Practitioner (AIF-C01) exam pattern: choosing between retrieval-augmented generation and fine-tuning. This page walks through a representative practice question, then explains the decision rule that covers most exam variations.
Short answer
The correct answer is B. Build a RAG pipeline using Amazon Bedrock Knowledge Bases.
Retrieval-augmented generation (RAG) is the correct fit because the source data changes weekly and answers must reflect the latest version. With Amazon Bedrock Knowledge Bases, you point the service at an Amazon S3 bucket of PDFs, Bedrock handles chunking, embedding, and storage in a managed vector store (Amazon OpenSearch Serverless, Aurora pgvector, Pinecone, or Redis), and the RetrieveAndGenerate API returns grounded answers with citations to the source documents. When the compliance team uploads a revised policy, you re-sync the data source and the new content is immediately retrievable — no retraining required. The decisive clues in the scenario are weekly updates, mandatory citations, limited ML expertise, and minimizing training effort. Fine-tuning, by contrast, bakes knowledge into model weights at training time; any new policy would require another customization job, a new provisioned throughput commitment, and validation. Fine-tuning also does not produce citations, which violates the compliance requirement. RAG solves freshness, traceability, and cost-per-update in a single managed pattern, which is why AIF-C01 scenarios with phrases like frequently updated documents, cite sources, or proprietary knowledge base almost always point to Bedrock Knowledge Bases rather than fine-tuning.
The Question
A financial services company wants a chatbot that answers employee questions using an internal policy library of roughly 4,000 PDF documents. The policies are revised every week by the compliance team, and answers must always cite the latest version. The team has limited ML expertise and wants to minimize training effort and ongoing maintenance. Which approach on AWS best fits this requirement?
Why B is correct
Retrieval-augmented generation (RAG) is the correct fit because the source data changes weekly and answers must reflect the latest version. With Amazon Bedrock Knowledge Bases, you point the service at an Amazon S3 bucket of PDFs, Bedrock handles chunking, embedding, and storage in a managed vector store (Amazon OpenSearch Serverless, Aurora pgvector, Pinecone, or Redis), and the RetrieveAndGenerate API returns grounded answers with citations to the source documents. When the compliance team uploads a revised policy, you re-sync the data source and the new content is immediately retrievable — no retraining required. The decisive clues in the scenario are weekly updates, mandatory citations, limited ML expertise, and minimizing training effort. Fine-tuning, by contrast, bakes knowledge into model weights at training time; any new policy would require another customization job, a new provisioned throughput commitment, and validation. Fine-tuning also does not produce citations, which violates the compliance requirement. RAG solves freshness, traceability, and cost-per-update in a single managed pattern, which is why AIF-C01 scenarios with phrases like frequently updated documents, cite sources, or proprietary knowledge base almost always point to Bedrock Knowledge Bases rather than fine-tuning.
Why the other options are wrong
Fine-tuning with Bedrock model customization changes model weights using labeled examples. It is expensive, slow, and must be re-run every time the policies change. It also does not produce citations back to source documents, which breaks the compliance requirement.
Continued pre-training adapts a base model to a domain vocabulary using large unlabeled corpora. It is the heaviest customization option, requires a provisioned throughput commitment on Bedrock, and still bakes stale knowledge into weights — the opposite of weekly freshness with low maintenance.
SageMaker JumpStart fine-tuning requires ML expertise, instance selection, training scripts, and an endpoint to host the customized model. The scenario explicitly calls for minimal training effort and limited ML skills, and fine-tuning still fails the freshness and citation requirements.
When to choose RAG vs fine-tuning on AWS
RAG and fine-tuning solve different problems, and the AIF-C01 exam tests whether you can tell them apart from scenario wording. You choose RAG when the goal is to ground a foundation model in external knowledge that it did not see during pre-training, especially when that knowledge changes often or must be cited. RAG leaves the model weights untouched. At query time, the system retrieves relevant chunks from a vector store and injects them into the prompt, so updates are as simple as re-indexing the source data. On AWS, the managed path is Amazon Bedrock Knowledge Bases, which wires together S3, an embedding model, a vector store, and the RetrieveAndGenerate API with citations built in. You choose fine-tuning when the goal is to change model behavior, tone, or format — for example, teaching a model a company-specific writing style, a structured output schema, or a narrow classification task. Fine-tuning bakes the pattern into weights, so inference is faster and prompts can be shorter, but every knowledge refresh means another training job. On AWS, fine-tuning is available through Amazon Bedrock model customization for supported foundation models (Titan, Llama, Claude Haiku on selected models) and through Amazon SageMaker JumpStart for open-source models with full control. Bedrock also offers continued pre-training for heavy domain adaptation on unlabeled corpora. The trade-offs: RAG wins on freshness, auditability, and setup cost, but adds retrieval latency and token cost per query. Fine-tuning wins on latency and prompt brevity, but costs more upfront and goes stale fast. Many production systems combine both — fine-tune for style and format, RAG for facts.
Ready to see how you'd score?
Take the free practice quiz and find out which AWS Certified AI Practitioner domains you need to focus on. No signup required.
Practice 5 similar questions
Same cert, same or adjacent domain. Use these after reviewing the explanation.