Foundation Model Integration, Data Management, and Compliance · Topic 1 of 9

Architecture Design, Proof of Concept and Standard Components

Design GenAI architectures for business needs, prove them with an Amazon Bedrock proof of concept, and standardize with the Well-Architected GenAI Lens.

  • 6 min
In a nutshell
  • Start from the constraints: data freshness, latency, cost, compliance and operational overhead decide the pattern.
  • RAG for private or changing knowledge, fine tuning for style and domain behavior, agents for multi step tasks.
  • Amazon Bedrock is serverless and multi model; SageMaker AI gives full control of hosting.

What it is

Every generative AI project starts with three decisions: which foundation model (FM) fits the job, how the application talks to it, and where it runs. The exam gives you a business goal plus constraints such as latency, data residency, budget or team skills, and asks for the design that satisfies all of them.

Before a full rollout, a proof of concept (PoC) checks that the idea works, performs well enough and is worth the money. Once it works, standard components and the AWS Well-Architected Generative AI Lens make sure every team builds it the same, safe way.

Business needgoal and constraintsshapesDesignFM, pattern, hostingvalidated byProof of conceptAmazon Bedrockapproved intoStandard componentstemplates and Lens review

How it works

Mapping requirements to design choices

Read the constraints first, then pick the pattern. Most scenarios fit one of a few integration patterns.

Requirement in the scenarioDesign choice
Answers must use private, frequently changing documentsRetrieval Augmented Generation (RAG) with Amazon Bedrock Knowledge Bases
Model must learn a style, format or domain vocabularyFine tuning or another customization, after prompting and RAG fall short
Task needs several steps and calls to business systemsAgent pattern with Amazon Bedrock Agents or Amazon Bedrock AgentCore
Simple text task with no private dataDirect model call with a well designed prompt
Many requests where no one waits for the answerAsynchronous or batch processing, such as Bedrock batch inference
Employees want a ready made assistant over company dataAmazon Q Business instead of a custom build

Deployment strategy is the second axis. Amazon Bedrock gives serverless access to many FMs through one API with no infrastructure to manage. Amazon SageMaker AI hosts models on endpoints you size and control, which suits open weight models, custom containers or special hardware needs.

Exam tip

"LEAST operational overhead" almost always points to Amazon Bedrock. Choose SageMaker AI only when the scenario needs something Bedrock cannot do, such as a model Bedrock does not offer or full control of the serving stack.

Building a proof of concept on Amazon Bedrock

A good PoC is small, measured and cheap to throw away. Amazon Bedrock suits it because you can try several models with no servers, then keep the winner.

Success criteriaquality, latency, costtest inBedrock playgroundseveral FMsmeasure withModel evaluationtest datasetcompareDecisionscale, change, stop
  1. Define success criteria up front. Name the accuracy, latency and cost per request that would justify production.
  2. Prototype quickly. Use the Bedrock console playgrounds and the Converse API, which uses one request format across models, so swapping models is easy.
  3. Measure with real data. Amazon Bedrock model evaluation runs automatic evaluations, LLM as a judge evaluations or human evaluations against your own prompt dataset.
  4. Check performance and cost. Record latency and token counts per request, then project the monthly cost at the expected volume.
  5. Decide on evidence. Present quality, latency and cost against the success criteria to the business owner.

Turn on Bedrock model invocation logging during the PoC. It records prompts and responses to Amazon CloudWatch Logs or Amazon S3 as evidence and a baseline.

Standardized technical components

When several teams build generative AI applications, each one should not reinvent prompt handling, logging or guardrails. Standard components make implementations consistent across deployment scenarios.

  • Reusable infrastructure as code. AWS CloudFormation or AWS CDK templates that create a knowledge base, a guardrail, logging and IAM roles the same way every time.
  • Shared libraries and services. One internal gateway or SDK wrapper for model calls, so retries, logging and guardrails are applied everywhere.
  • Managed shared assets. Prompts in Amazon Bedrock Prompt Management and guardrails in Amazon Bedrock Guardrails, referenced by ID and version instead of copied into code.

The AWS Well-Architected Framework has six pillars: operational excellence, security, reliability, performance efficiency, cost optimization and sustainability. The Generative AI Lens applies those pillars to generative AI workloads, with questions about model selection, prompt and data management, responsible AI and monitoring.

You can apply the Generative AI Lens in the AWS Well-Architected Tool (AWS WA Tool) to review a workload. The review records risks and improvement plans, and teams can repeat it at each milestone to keep designs consistent.

NeedUse
Repeatable infrastructure for every teamCloudFormation or CDK templates
Review a generative AI workload against best practicesGenerative AI Lens in the AWS WA Tool
One version controlled prompt used by many appsAmazon Bedrock Prompt Management
One safety policy applied to many appsAmazon Bedrock Guardrails

How the exam asks about this

  • "A company wants to test three FMs on its own support tickets before committing, with minimal infrastructure" -> Amazon Bedrock model evaluation with a custom prompt dataset.
  • "Ten teams build chat assistants and leadership wants a consistent best practice review" -> Generative AI Lens in the AWS Well-Architected Tool.

Common mistakes

Fine tuning to add facts. Fine tuning teaches style and patterns, but facts that change need RAG. Retraining every week is slow and expensive.

A PoC with no success criteria. Without agreed thresholds, a demo that looks good cannot justify production. Define quality, latency and cost targets before building.

Choosing SageMaker AI by default. Self hosting adds endpoint sizing, scaling and patching. If Bedrock offers a suitable model, it is the lower overhead answer.

Watch out

A PoC measured on a handful of hand picked prompts overstates quality. Use a representative dataset that includes hard and edge cases.

What to remember for the exam

  • Start from the constraints: data freshness, latency, cost, compliance and operational overhead decide the pattern.
  • RAG for private or changing knowledge, fine tuning for style and domain behavior, agents for multi step tasks.
  • Amazon Bedrock is serverless and multi model; SageMaker AI gives full control of hosting.
  • A PoC needs success criteria, real test data, Bedrock model evaluation and cost projection.
  • The Converse API keeps one request format across models, which makes comparing models easy.
  • The Generative AI Lens in the AWS WA Tool reviews workloads against the six Well-Architected pillars.

Check yourself

  • A team must validate that an FM can summarize claims accurately before a full build. What is the lowest effort approach?

    Prototype on Amazon Bedrock and run a model evaluation job against a representative dataset of real claims, measured against agreed success criteria.

  • How do you give many teams a consistent review of their generative AI workloads?

    Apply the Generative AI Lens in the AWS Well-Architected Tool and track the risks and improvement plans it records.

  • Product data changes daily and answers must reflect it. Fine tuning or RAG?

    RAG, because retrieval reads current data at query time while fine tuning freezes knowledge at training time.

Topic 1 of 9 in Foundation Model Integration, Data Management, and Compliance

Test yourself on Foundation Model Integration, Data Management, and Compliance

The full AWS AIP-C01 question bank has 827 questions, with timed practice exams and an explanation for every answer.

This study guide is independent and is not affiliated with, endorsed by, or sponsored by AWS. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.