Day 2 - Domain Review 2: Foundation Model Applications (AWS AI Services) Key Summary

📌 Key Takeaways

• AWS AI services come in three layers: (1) ready-to-use AI services (APIs), (2) SageMaker for building models yourself, and (3) Bedrock and Amazon Q for generative AI.
• Even with the same data, the verb separates the services. Pulling out the text is Textract, analyzing the meaning is Comprehend, creating something new is Bedrock.
• "Just calling an API = AI service / borrowing an FM = Bedrock / building and controlling the model yourself = SageMaker."
• The ways to adapt an FM to your use get more expensive in this order: prompt → RAG → fine-tuning → training from scratch. Up-to-date and in-house knowledge calls for RAG; internalizing a tone and expertise calls for fine-tuning.
• FM evaluation is done with human evaluation, benchmark datasets, automatic metrics (ROUGE, BLEU, BERTScore), and a judge model.

Layer 1: Ready-to-Use AI Services (Pretrained APIs)

Today we organize Domain 3 (foundation model applications, 28%), which carries the most weight, and the service selection that shows up across several domains. AIF-C01 endlessly asks "Which AWS service best fits this situation?" The ability to quickly recall pairs of a service name and its "one-line purpose" is the score itself.

The first layer is managed services you use just by calling an API, with no ML knowledge. Each does "one thing" well.

What these services have in common is that "you borrow a model AWS has already trained." Users just call the API without worrying about data, infrastructure, or the training process. So when "a team without ML specialists wants to add a feature quickly" appears, a service from this layer is almost always the answer. Conversely, if "we must build a model that fits our own data exactly," you move on to the next layer (SageMaker).

• The list of in-scope services in the current exam guide (machine learning category) includes Bedrock, Bedrock AgentCore, Comprehend, Lex, Nova, Personalize, Polly, Rekognition, SageMaker AI, SageMaker JumpStart, Textract, Transcribe, Translate, and AWS Transform. The list explicitly states that it is "not exhaustive and may change," so even for Forecast and Kendra, which are not on the list, know their roles, but spend more review time on the services that are on the list.
• AWS documentation notes that Amazon Forecast is no longer available to new customers, while existing customers can keep using it. The pairing "time-series forecasting = Forecast" is still a clue for solving questions.

Pairs That Get Confused Within Layer 1

• Services are often chained together. Call recording → Transcribe → Comprehend (sentiment); scanned document → Textract → Comprehend (entities); foreign-language speech → Transcribe → Translate → Polly. When a combination appears in the answer choices, check that the order of the arrows matches the changes in the form of the data (speech → text → analysis).

Layer 2: SageMaker, for Building Models Yourself

Amazon SageMaker is a fully managed platform for building, training, and deploying ML models yourself. If the API services above are "off-the-shelf," SageMaker is a "custom workshop." It covers the entire ML lifecycle from data preparation through training, tuning, deployment, and monitoring. The current AWS documentation and exam guide write this service under the name Amazon SageMaker AI.

The key distinction: "just calling an API = AI services layer / building and controlling the model yourself = SageMaker."

What SageMaker takes responsibility for is "automating the repetitive work of the ML lifecycle." Model training comes with endless data cleaning, algorithm selection, hyperparameter tuning, GPU infrastructure management, and deployment and monitoring. SageMaker bundles these stages as a managed service so data scientists can focus on the model itself. So when you see words like "custom model," "train it yourself," or "MLOps," think of SageMaker.

• JumpStart is easy to confuse between Bedrock and SageMaker. Both handle pretrained models, but the difference is that with JumpStart, you host, fine-tune, and control it yourself in the SageMaker environment.

Layer 3: Generative AI — Bedrock and Amazon Q

Amazon Bedrock is a fully managed service that lets you call foundation models from multiple companies through a single API. You can use a variety of FMs, such as those from Anthropic, Meta, and Amazon (Titan, Nova), without managing infrastructure, and you can customize them with your own data or set up RAG and agents. If you want to "host and serve FMs" or "pick from several models," it is Bedrock.

Amazon Q is AWS's generative AI assistant. There are two branches.

When you set up RAG yourself on top of Bedrock, two things often appear together: Knowledge Bases for Bedrock (embeds documents and provides them as grounding for search) and Agents for Bedrock (automatically performs multi-step tasks). Their names in current documentation are Amazon Bedrock Knowledge Bases and Amazon Bedrock Agents, respectively. The safety feature likewise circulates under both its old name "Guardrails for Amazon Bedrock" and its current name Amazon Bedrock Guardrails. Whichever way it appears in a question, it is the same feature.

💡 Related theory: The relationship between Bedrock and Amazon Q is closer to "engine and finished product." Bedrock is a building block for taking an FM and building an app yourself, and Amazon Q is a product already completed on top of it in the form of an assistant. So if "a developer builds a generative AI app themselves," Bedrock is the answer, and if "a ready-to-use in-house AI assistant is needed," Amazon Q is the answer. The decisive difference from SageMaker is that both use off-the-shelf FMs without training a new model.

• The current exam guide gives, as examples of AWS services for developing generative AI applications, Bedrock, SageMaker AI, and SageMaker JumpStart along with Amazon Quick (a work AI helper that supports research, business insights, and automation), Kiro (an AI-based development environment), Strands Agents (an open-source framework for building agents), and Amazon Bedrock AgentCore (a platform for securely deploying and operating agents built with any framework or model). Since this is an introductory exam, it is enough to be able to pair each name with a one-line purpose.

Five Ways to Adapt an FM to Your Use

The central question of Domain 3 is "What do you use to adapt the model to our work, and at what cost?"

• Model distillation goes in a different direction. It teaches a small, cheap model (the student) with the responses of a large, expensive model (the teacher), aiming to get similar quality at lower cost and latency. The clue is "we want to lower inference cost while keeping quality."
• Fine-tuning data needs curation, governance, sufficient size, accurate labels, and representativeness. A method that uses human evaluation as a reward signal to refine the model (RLHF) is also a term in the exam guide.
• Branches of fine-tuning include instruction tuning, which refines the model to fit an instruction format; domain adaptation to a specific domain; and transfer learning, which carries what was learned on one task over to another.

The Parts That Make Up RAG

• Flow: split documents into chunks → an embedding model turns the chunks into vectors → store them in a vector database → turn the question into a vector as well and find the nearest chunks → put the found chunks into the prompt and have the FM answer.
• The AWS services the exam guide gives as vector databases that store embeddings are Amazon OpenSearch Service, Amazon Aurora, Amazon Neptune, and Amazon RDS for PostgreSQL. The list of in-scope services also includes Amazon DocumentDB.
• Amazon Bedrock Knowledge Bases bundles this flow (chunking, embedding, storing, searching) as a managed service.
• RAG does not change the model. If you edit a document, it is reflected from the very next question. This is why RAG suits frequently changing knowledge.

Prompt Engineering: Techniques and Risks

• As components of a prompt, the exam guide lists context, instruction, and negative prompt. The principles of a good prompt are to be specific yet concise, and to refine it through experimentation.

Choosing and Evaluating Models

• Selection criteria: cost, modality (text, images, etc.), latency, multilingual support, model size and complexity, whether it can be customized, input and output length, prompt caching.
• Inference parameters: questions ask about the effect of temperature and input/output length on the response (summarized yesterday).
• Pricing methods: there is pay-per-token pricing, and provisioned throughput, which reserves processing capacity in advance. If usage is uneven, the former; if steady, high throughput must be guaranteed, the latter.

• A high metric is useless if it does not fit the business goal. The exam guide gives task completion rate, user satisfaction, and cost per interaction as metrics for alignment with business goals.

Quick Selection Guide: Situation → Service

The exam is ultimately a "situation → service" mapping. Here are the representative patterns, condensed.

• Extract text and tables from scanned receipts/contracts → Textract
• Identify objects, faces, and inappropriate content in photos → Rekognition
• Analyze positive/negative sentiment in customer reviews → Comprehend
• Call center recordings to text → Transcribe; the reverse → Polly
• Build a customer service chatbot → Lex (or Q Business)
• Product recommendations for an online shop → Personalize
• Pick a multilingual FM and integrate it into an app → Bedrock
• Train and deploy a custom model yourself → SageMaker

💡 Related theory: Even with the same "text," a different task means a different service. "Pulling out" the text is Textract (document OCR), "analyzing the meaning" of the text is Comprehend (NLU), and "creating" text is Bedrock/an LLM. The exam traps hide in these subtle verb differences ("extract vs. analyze vs. generate"), so building the habit of catching the verb in the question first raises your accuracy a great deal.

Three Questions to Decide Which Layer

① Is one ready-made function enough? (dictation, translation, OCR, recommendation...)
   └─ Yes ─► Layer 1 AI services

② Do you need to create new content or hold free-form conversation?
   ├─ A ready-to-use assistant is needed ─► Amazon Q (work = Q Business, coding = Q Developer)
   └─ You build the app yourself ──────────► Amazon Bedrock

③ Must you train, tune, deploy, and control the model itself with your own data?
   └─ Yes ─► Amazon SageMaker (quick start with pretrained models = JumpStart)

• Look from the layer with the least operational burden first. If ① works, ①; if not, ②; if that still does not work, ③. When a qualifier such as "with minimal ML expertise" is attached, ③ is almost always the wrong answer.

In the next piece, we will summarize responsible AI and security/governance, corresponding to Domains 4 and 5, focusing on the main text. "Safe and responsible" carries as much weight on the exam as the technology does.

📖 Glossary

• OCR: Technology that reads the text in images or scanned documents as text.
• STT / TTS: Converting speech to text / converting text to speech.
• Intent: The user's intention that a chatbot must identify. Example: "cancel an order."
• RAG (Retrieval-Augmented Generation): A method that retrieves relevant documents, puts them into the prompt, and has the model answer based on them.
• Vector database: A store that holds embedding vectors and finds those that are close in meaning.
• Fine-tuning: Additionally training a pretrained model on our own data to change its weights.
• Model distillation: A method of teaching a small model with the responses of a large model to lower cost and latency.
• chain-of-thought: A prompt technique that has the model write out its reasoning step by step before giving an answer.
• Provisioned throughput: A pricing method where a fixed processing capacity is reserved in advance and used.
• LLM-as-a-judge: An evaluation method where one LLM grades the responses of another LLM.

📝 Practice Questions

Click a choice to reveal the answer and explanation.

Question 1

Auto-extract text, tables, input form fields from scanned paper contract. Most fitting service?

• AAmazon Comprehend
• BAmazon Textract
• CAmazon Polly
• DAmazon Translate

Question 2

Without ML experts, want to call several companies' foundation models via single API and build generative AI app. Most fitting service?

• AAmazon SageMaker
• BAmazon Rekognition
• CAmazon Bedrock
• DAmazon Forecast

Question 3

Data science team wants to directly train custom ML model with company data, tune it, deploy it, and post-deployment monitor drift. Most fitting?

• AAmazon SageMaker
• BAmazon Q Business
• CAmazon Lex
• DAmazon Kendra

Question 4

Build customer voice/chat bot with intent recognition and conversation flow management specialization. What service?

• AAmazon Polly
• BAmazon Transcribe
• CAmazon Lex
• DAmazon Personalize

Question 5

Critical difference between Bedrock and SageMaker?

• ABedrock uses ready-made foundation models, SageMaker builds custom models from scratch·training directly
• BIdentical, names only differ
• CBedrock image-only, SageMaker text-only
• DSageMaker can't train models