cert-bank

Guides

이 글을 한국어로 읽기

Prompting, RAG, Fine-Tuning or Continued Pre-Training: One Rule That Works on AWS, Azure, Google Cloud and Databricks Exams

Every generative AI exam asks some version of the same question: a model does not do what the business wants, so what do you change? The options are always the same four. Write a better prompt. Give the model documents at question time (retrieval-augmented generation, RAG). Train it further on labelled examples (fine-tuning). Or train it further on large amounts of unlabelled text from your field (continued pre-training). Across five exams (AWS AIF-C01 and AIP-C01, Microsoft AI-103, Databricks DBX-GAI and Google GCP-GAIL), 131 of our practice questions mention RAG and 101 mention fine-tuning in the question or its options. The vendors use different product names, but the decision rule is the same everywhere.

Concept comparison · AIF-C01, AIP-C01, AI-103, DBX-GAI, GCP-GAIL and other generative AI exams

Prompting, RAG, Fine-Tuning or Continued Pre-Training: One Rule That Works on AWS, Azure, Google Cloud and Databricks Exams

The rule in one sentence

If the model is missing knowledge, especially knowledge that changes or must be cited, retrieve it. If the model is missing behaviour (a tone, a format, a task pattern) and you have labelled examples, fine-tune. If it lacks the language of a whole field and you have plenty of unlabelled domain text, continue pre-training. Otherwise, stay with prompting. On every exam, try the cheapest rung first and climb only when the stem gives a reason.

TechniqueWhat it changesWhat it needsClue in the stem
Prompt engineeringOnly this request: instructions, a few examples, a formatNothing but wording"quickly", "no training data", "weights cannot change"
RAG / groundingThe context the model reads at question timeDocuments, an index and a retriever"changes daily", "internal documents", "cite sources", "remove content quickly"
Fine-tuningThe weights, toward a behaviourLabelled prompt and response pairs"brand voice", "always this format", "labelled examples"
Continued pre-trainingThe weights, toward a domain's languageLarge amounts of unlabelled domain text"millions of unlabelled reports", "domain vocabulary"

Cost and effort rise from top to bottom. Training a model from scratch sits above all four. It needs huge data and compute, and it is almost never the answer.

Beginner tip: before you read the options, underline the noun the stem says is missing, and write K (knowledge) or B (behaviour) next to it. "Answers do not reflect this week's price list" is K. "Replies must always use our three-part support layout" is B. Then check whether the stem mentions labelled pairs or unlabelled text. Two marks on the question paper usually remove two of the four options.

Prompt, RAG, fine-tuning or continued pre-training?
Is knowledge missing, especially facts that change or must be cited?
YES →
RAG: retrieve it at question time
↓ no
Is behaviour missing (a tone, a format, a task pattern), with labelled examples?
YES →
Fine-tuning
↓ no
Does it lack a whole field's language, with plenty of unlabelled domain text?
YES →
Continued pre-training
↓ no
OTHERWISE →
Stay with prompt engineering, the cheapest rung
Mark what the stem says is missing, then take the first yes; climb past prompting only when the stem gives a reason.

The same rule in each vendor's words

AWS (AIF-C01, AIP-C01)

More of our questions on this distinction come from the AWS exams than from any other vendor. AIF-C01 asks for the cost tradeoffs of the customization approaches (in-context learning, RAG, fine-tuning, pre-training, distillation), which works as a cost ladder: start with the cheapest step that works. AIP-C01 adds the operating details. A Bedrock model you customize becomes a private custom model. It usually needs dedicated capacity (Provisioned Throughput) rather than the shared on-demand pool; for a few base models, a custom model deployment also allows on-demand inference. Both exams also separate two look-alikes. Distillation trains a smaller model to imitate a larger one, to save cost and latency. Prompt caching handles repeated identical prompt prefixes. Neither one adds knowledge.

Google Cloud (GCP-GAIL)

GCP-GAIL uses the same staircase: start with prompting, ground the model in trusted data when answers must reflect company facts, and tune only when prompting cannot fix recurring errors. GCP-GAIL adds a split inside grounding. Public, fast-changing world facts point to Grounding with Google Search. Your own documents point to RAG over your data (pre-built RAG with Agent Search, or the RAG APIs). As a Generative AI Leader exam, it also asks about data terms: Google Cloud states that it does not use customer data to train or fine-tune models without the customer's permission or instruction.

Microsoft Azure (AI-103)

AI-103's skills outline asks for RAG, grounding and knowledge integration, and does not list fine-tuning at all. Microsoft's Foundry documentation says fine-tuning "doesn't replace retrieval for current information". Our AI-103 questions build on that: a tuned model cannot cite sources, and it must be trained again whenever the content changes. Knowledge comes from tools instead. The file search tool searches documents you or your users upload. The Azure AI Search tool connects to an existing index. A Foundry IQ knowledge base can be shared by several agents (some of its features are still in preview). Web search, which uses Grounding with Bing Search, covers current public information. In our questions, fine-tuning appears only as a later step for consistent style and format, after prompting and structured output have been tried.

Databricks (DBX-GAI)

DBX-GAI assumes retrieval for private knowledge that changes, because passages are read at question time while tuned weights stay a snapshot. Most of its questions are about doing RAG well: parsing, chunking, AI Search indexes and judges. Fine-tuning appears on the operations side. A fine-tuned model of a supported architecture is registered in Unity Catalog and served with provisioned throughput. Before any tuning, the licence of the base model and the training data must permit the use. A non-commercial licence rules out a paid product.

Going deeper: why weights are a poor place to keep facts

Fine-tuning adjusts many weights a little, so that certain outputs become more likely. A fact learned this way is not stored as a record you can find, update or delete. It is spread across the network and mixed with everything else the model learned. That has three consequences the exams build on. The model cannot point to where an answer came from, so it cannot cite. Changing one fact means training again, and the old version may still surface. Removing content (an expired licence, a deletion request) has no clean operation. Retrieval keeps facts as rows and documents: update the source, re-sync the index, and the next answer reads the new text. This is why "must cite" and "must be removable" point to RAG even when fine-tuning could, in principle, teach the same facts.

Traps that appear on more than one exam

  • Fine-tuning to teach new facts. It is a common wrong answer on every vendor's exam. The tuned model goes stale as soon as the data changes.
  • Fixing bad retrieval with generation settings or fine-tuning. If the retriever returns the wrong passages, fix chunking, embeddings or filters. A fluent model cannot repair missing context.
  • Too little fine-tuning data. A small set leads to memorisation and overfitting. Quality (clean, relevant, consistent, deduplicated examples) matters more than volume.
  • Personal data in training sets. A model can memorise and repeat what it was trained on. Remove or mask personal data before training. Runtime guardrails are only a second layer.
  • Confusing the training-based techniques. Labelled pairs mean fine-tuning. Unlabelled domain text means continued pre-training. Human preference rankings with a reward model mean RLHF. A small student copying a large teacher means distillation.
  • Thinking it must be either-or. A fine-tuned model for a consistent voice, backed by RAG for current facts, is a legitimate combined answer when the stem asks for both.
Going deeper: what separates continued pre-training from fine-tuning

Both update the weights. The difference is the training signal. Continued pre-training uses the same objective as the original pre-training: predict the next token of ordinary text. It needs no labels, so it can use a large pile of a field's documents, and it shifts the model's vocabulary and background knowledge toward that field. Fine-tuning uses pairs of an input and the desired output. It teaches the model to respond in a particular way, and every pair has to be prepared by someone. That is why the data in the stem decides the answer: "large volumes of unlabelled clinical notes" means continued pre-training, while "two thousand reviewed question-and-answer pairs" means fine-tuning.

Practise

Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.

Counts come from our own question bank. Product names and customisation options differ by vendor and change often; check each exam's official guide and the vendor's current documentation.