Amazon Bedrock is the centerpiece of AWS’s generative AI strategy and the most-tested service on the AWS Certified AI Practitioner (AIF-C01) exam. If you understand Bedrock deeply — what it is, what it isn’t, and how its features fit together — a substantial portion of the exam becomes mechanical.
This guide is built specifically for AIF-C01 candidates. It covers exactly what AWS tests about Bedrock, the scenarios where Bedrock is and isn’t the right answer, and the patterns that consistently trip up candidates.
What Is Amazon Bedrock?
Amazon Bedrock is a fully managed service that provides API access to foundation models from multiple providers, plus a layered set of features for building generative AI applications without managing any infrastructure.
The key phrases for the exam:
- Fully managed — no infrastructure to provision or maintain
- Serverless API — call models via REST or SDK
- Multiple model providers — single API for access to many foundation model families
- Application building blocks — knowledge bases, agents, guardrails, evaluation, model customization
If a question asks for a way to use foundation models with no infrastructure, no deployment overhead, and access to multiple model providers, the answer is almost always Bedrock.
Why Bedrock Matters for AIF-C01
Bedrock shows up in questions across all five AIF-C01 domains:
- Domain 2 (Generative AI Fundamentals): What Bedrock is and how it differs from SageMaker
- Domain 3 (Applications of Foundation Models): Bedrock features for prompt engineering, RAG, and customization
- Domain 4 (Responsible AI): Bedrock Guardrails for content filtering
- Domain 5 (Security): Securing Bedrock with IAM, VPC endpoints, and PrivateLink
Realistically, 6 to 12 questions on your exam will touch Bedrock in some form.
Core Bedrock Concepts You Must Know
Foundation Models on Bedrock
Bedrock provides access to a curated catalog of foundation models from multiple providers, including text generation, embedding, image generation, and multimodal models. For AIF-C01 you don’t need to memorize specific model names or version numbers — but you do need to know:
- Different models have different strengths (some optimized for reasoning, some for cost, some for speed, some for multimodal inputs)
- Models have context window limits (the amount of text the model can consider)
- Different models have different pricing, typically per input/output token
- You can switch models with minimal code changes since they’re behind a unified API
Bedrock Inference
You invoke a Bedrock model by sending a prompt and inference parameters. Common parameters tested on the exam:
- Temperature — controls randomness. Higher = more diverse and creative. Lower = more deterministic.
- Top-p (nucleus sampling) — controls diversity by sampling from the smallest set of tokens whose cumulative probability exceeds p.
- Top-k — restricts sampling to the top k most likely tokens.
- Max tokens — caps the response length.
- Stop sequences — strings that, when generated, stop the response.
Exam pattern: “The output is too repetitive — what should you change?” → raise temperature. “The output is too random — what should you change?” → lower temperature.
Bedrock Knowledge Bases (Managed RAG)
Bedrock Knowledge Bases is AWS’s managed implementation of Retrieval-Augmented Generation (RAG). It does the heavy lifting:
- Ingest documents from S3
- Chunk them
- Generate embeddings automatically
- Store them in a vector database (Bedrock-managed, OpenSearch, Aurora, Pinecone, etc.)
- Retrieve relevant chunks at query time and inject them into the model’s context
When to use Knowledge Bases:
- You need responses grounded in your private documents
- Your documents change frequently and you don’t want to retrain
- You need citations or source references in responses
When not to use Knowledge Bases:
- You want to teach a model a stable style or behavior (use fine-tuning instead)
- Your data fits comfortably in a single prompt’s context (use prompt engineering)
- You need the model to perform actions, not just answer questions (use Agents)
Bedrock Agents
Bedrock Agents extend foundation models with the ability to take actions — calling APIs, querying data sources, executing workflows. An Agent typically:
- Receives a user request
- Plans the steps needed (using a foundation model for reasoning)
- Calls the relevant action groups (Lambda functions or API integrations)
- Returns a response that combines the action outputs
When to use Agents:
- The user request requires looking something up and doing something (e.g., “check my balance and transfer $100”)
- The model needs to interact with external systems
- You need multi-step reasoning with tool use
Bedrock Guardrails
Bedrock Guardrails is the responsible AI layer for generative outputs. It enables:
- Denied topics (e.g., medical advice, legal advice)
- Content filters (hate, insults, sexual, violence)
- PII detection and redaction in inputs and outputs
- Word filters for custom blocked terms
- Contextual grounding checks to reduce hallucinations
When to use Guardrails:
- Customer-facing applications where harmful or off-topic outputs are unacceptable
- Regulated industries (healthcare, finance) requiring strict content controls
- Any production deployment of a foundation model
Exam trap: Don’t confuse Bedrock Guardrails with SageMaker Clarify. Guardrails = generative content filtering. Clarify = bias and explainability for classical ML.
Bedrock Model Customization
Bedrock supports several customization options. AIF-C01 tests your ability to choose the right one:
| Option | Use When… |
|---|---|
| Prompt engineering | Behavior adjustment with no training. Fastest, cheapest. |
| Knowledge Bases (RAG) | You need responses grounded in changing private data. |
| Fine-tuning | You have labeled examples and want stable style/behavior changes. |
| Continued pre-training | You have large amounts of unlabeled domain text and want the model to absorb domain knowledge. |
The decision rule of thumb:
- Fast and free? → Prompt engineering
- Changing facts? → RAG
- Stable style or behavior? → Fine-tuning
- Domain-specific language at scale? → Continued pre-training
Bedrock vs. Other AWS AI Services
This comparison table is one of the most exam-relevant artifacts you can build:
| Service | Use For | Don’t Use For |
|---|---|---|
| Amazon Bedrock | Calling foundation models via API, building generative AI apps | Training custom ML models from scratch |
| Amazon SageMaker | Building, training, and deploying custom ML models end-to-end | Quick API access to existing foundation models |
| SageMaker JumpStart | Browsing and deploying pre-trained models (incl. foundation models) to your own SageMaker endpoints | Fully managed serverless inference |
| Amazon Q Developer | AI assistant for software development (IDE, code) | General enterprise knowledge |
| Amazon Q Business | AI assistant connected to enterprise data sources | Code generation in IDE |
| Amazon Comprehend | Pre-built NLP tasks (sentiment, entities, PII) | Generative text creation |
| Amazon Rekognition | Pre-built computer vision tasks | Custom multimodal generative AI |
A common AIF-C01 question pattern: “Choose the AWS service for [specific generative AI use case] with [specific constraint].” The right answer is usually Bedrock unless the constraint pushes you elsewhere (e.g., “the team needs full control over training” → SageMaker; “the team is building an IDE assistant” → Q Developer).
Securing Bedrock for the Exam
Security questions about Bedrock test your knowledge of standard AWS security primitives applied to generative AI:
- IAM — controls who can invoke models and access knowledge bases
- KMS encryption — for data at rest in knowledge bases and customization jobs
- TLS — for data in transit
- VPC endpoints (PrivateLink) for Bedrock — keeps traffic on the AWS backbone, never traversing the public internet
- CloudTrail — logs all Bedrock API activity (who invoked what, when)
- Bedrock model invocation logging — captures model inputs and outputs for audit
- Resource-based policies for fine-grained access controls
Exam pattern: “All traffic to Bedrock must stay within the AWS network.” → VPC endpoint via PrivateLink.
Common Bedrock Question Patterns on AIF-C01
After analyzing many practice questions, a few patterns repeat:
Pattern 1: “Choose Bedrock vs. SageMaker JumpStart”
Both can host foundation models, but the differentiators:
- Bedrock is a fully managed serverless API with built-in features (knowledge bases, agents, guardrails). No endpoints to deploy.
- SageMaker JumpStart provides pre-trained models you typically deploy to SageMaker endpoints. More control, more responsibility.
If the question emphasizes “no infrastructure,” “fastest to implement,” or “no ML expertise,” choose Bedrock. If it emphasizes “full control,” “custom training pipelines,” or “advanced ML features,” choose SageMaker.
Pattern 2: “RAG vs. Fine-Tuning”
When the use case involves answering questions from private, frequently changing data: RAG (Bedrock Knowledge Bases).
When the use case involves teaching the model a stable tone, format, or behavior: fine-tuning.
Pattern 3: “Choose the Inference Parameter”
Output too repetitive → raise temperature. Output too random → lower temperature. Output too long → lower max tokens. Output sometimes runs into other content → use stop sequences.
Pattern 4: “Choose the Responsible AI Tool”
Filter inappropriate or off-topic generative content → Bedrock Guardrails. Detect bias in classical ML predictions → SageMaker Clarify. Document model intended use and limitations → Model Cards or AI Service Cards.
Pattern 5: “Make Bedrock Traffic Private”
Keep Bedrock traffic off the public internet → VPC endpoint for Bedrock (PrivateLink).
How Bedrock Fits Into the AWS Generative AI Stack
A simplified mental model for the exam:
User
│
▼
Application
│
▼
Amazon Bedrock API
│
├─► Foundation Model (chosen at invocation time)
├─► Knowledge Base (optional RAG layer)
├─► Guardrails (optional content controls)
└─► Agents (optional action-taking layer)For most application architectures shown in exam questions, Bedrock sits as the foundation model access layer with optional features layered on top.
Bedrock Pricing Concepts
You don’t need to memorize prices, but you should know:
- Bedrock charges per token (input + output) for on-demand inference
- Provisioned Throughput is available for guaranteed capacity at a fixed cost — appropriate for high-volume production workloads
- Knowledge Bases charge for the underlying vector store and the embedding generation
- Fine-tuning has training costs plus storage
- Guardrails are typically billed per text policy evaluation
Exam pattern: “The team needs predictable cost and capacity for a high-volume production workload.” → Provisioned Throughput.
Quick Bedrock Reference Table
| Feature | What It Does | When to Use |
|---|---|---|
| Foundation models | Generate text, images, embeddings | Any generative AI use case |
| Knowledge Bases | Managed RAG | Grounded responses from private data |
| Agents | Take actions, multi-step reasoning | When the model needs to do things, not just answer |
| Guardrails | Content filtering | Production deployments, regulated industries |
| Fine-tuning | Custom model training | Stable style or behavior teaching |
| Continued pre-training | Domain knowledge absorption | Large unlabeled corpora |
| Model evaluation | Benchmark models | Choosing between model options |
| Provisioned Throughput | Reserved capacity | High-volume, predictable workloads |
| VPC endpoint | Private network access | Compliance, security requirements |
| Model invocation logging | Audit trail | Compliance, debugging |
Print this table. Memorize it. Most Bedrock questions on AIF-C01 boil down to a single row of this table.
Common Bedrock Mistakes on the Exam
- Choosing SageMaker when Bedrock is the answer. Bedrock’s “no infrastructure” position is its key differentiator.
- Choosing fine-tuning when RAG is the answer. RAG is the right pattern for changing facts.
- Choosing Guardrails for bias detection. Guardrails filter generative content; Clarify analyzes classical ML bias.
- Forgetting PrivateLink. Network security questions about Bedrock almost always have PrivateLink in the right answer.
- Conflating Bedrock and SageMaker JumpStart. Both can deliver foundation models, but Bedrock is serverless and JumpStart deploys to your own endpoints.
FAQ: Amazon Bedrock for AIF-C01
Q: Do I need to know specific model names for the exam? A: No. AIF-C01 focuses on Bedrock as a service and its features, not on individual model naming.
Q: How many Bedrock questions are on the exam? A: AWS doesn’t publish exact numbers, but 6 to 12 of the 65 questions typically touch Bedrock concepts.
Q: Are Knowledge Bases the same as RAG? A: Knowledge Bases are AWS’s managed implementation of RAG. The pattern is RAG; the AWS product is Knowledge Bases.
Q: When would I choose SageMaker over Bedrock? A: When you need full control over the training pipeline, want to host your own custom models, or are doing classical ML beyond foundation models.
Q: Are Bedrock Guardrails the same as Content Moderation in Comprehend? A: No. Guardrails are designed for generative AI inputs and outputs in Bedrock. Comprehend has separate moderation features for general text.
Q: How does Bedrock compare to building on EC2? A: Bedrock is fully managed and serverless. EC2 requires you to manage infrastructure, scaling, GPU drivers, model loading, and inference servers. AIF-C01 will always favor the managed option in scenarios that don’t require deep customization.
Conclusion
Amazon Bedrock is the most exam-relevant AWS service for AIF-C01. Master its core concepts — foundation model access, Knowledge Bases, Agents, Guardrails, and the customization options — and you’ll handle a substantial portion of the exam confidently.
The patterns to internalize are simple: Bedrock is the serverless foundation model API; Knowledge Bases handle changing facts (RAG); Agents handle action-taking; Guardrails handle content filtering; fine-tuning handles stable style or behavior. PrivateLink keeps traffic private, and IAM controls access.
Want to see how well you handle Bedrock scenarios under exam conditions? Try our AWS Certified AI Practitioner mock exam bundle — 8 full-length exams with detailed Bedrock coverage and explanations on every question, calibrated for the real AIF-C01 difficulty.