Knowledge bases (managed RAG) — AIF-C01

WHAT IT IS

Knowledge Bases for Amazon Bedrock is a managed service that implements Retrieval-Augmented Generation (RAG) — a technique that draws information from a private data store to augment the responses generated by a foundation model. Rather than requiring you to build and operate the retrieval pipeline yourself, Amazon Bedrock manages the underlying data ingestion, indexing, storage, and retrieval infrastructure so you can focus on your application and agent logic.

RAG itself is the process of optimizing the output of a large language model so it references an authoritative knowledge base outside of its training data sources before generating a response.

Mental model

Think of RAG as giving the model a live reference shelf. The model's training is fixed, but before it answers your question, it first searches the shelf for relevant pages, then reads those pages while writing its answer. Knowledge Bases for Amazon Bedrock automates the work of building, stocking, and searching that shelf.

Two distinct phases occur:

Ingestion phase (done once, then kept current): data is split into chunks, each chunk is converted to a numerical vector embedding by an embedding model, and those vectors are written to a vector index alongside a mapping back to the original document.

Runtime phase (every query): the user's query is also converted to a vector, the index is searched for chunks whose vectors are semantically similar to the query vector, and those chunks are added to the prompt before the foundation model generates its answer.

When to use it

The exam frequently tests the boundary between Knowledge Bases (managed RAG) and full model retraining. The decision turns on whether the model needs to reference new information or learn a new behavior.

Situation	Appropriate approach	Reason
Model needs access to your private or current documents it was not trained on	Knowledge Bases (RAG)	Retrieves at query time; no retraining needed
Model answers are outdated because training data is stale	Knowledge Bases (RAG)	External data can be updated asynchronously
Model needs to adopt a new writing style or follow domain conventions	Fine-tuning	Behavior is baked into weights, not retrieved
Model consistently hallucinates on a narrow factual domain despite retrieval	Fine-tuning or continued pre-training	Pattern recognition, not lookup, is needed
You want to cite sources and let users verify answers	Knowledge Bases (RAG)	Supports response citations back to original documents
You have no labeled training data but have a large document corpus	Knowledge Bases (RAG)	No labels required; documents are indexed as-is

COMMON MISCONCEPTION

"Using a knowledge base is the same as retraining or fine-tuning the model."

This is the specific trap the exam exploits. RAG does not change the foundation model's weights. The model's parameters remain frozen; it simply receives additional context through the prompt at inference time. Retraining updates what the model knows intrinsically; RAG updates what the model can look up before answering.

A second, closely related misconception: "RAG is only useful for question-answering chatbots." The official documentation describes Knowledge Bases as enabling applications that answer user queries, augment custom prompts with retrieved information, and support agentic AI applications with multi-hop reasoning — not just simple Q&A.

A third misconception: "You must build and manage the vector database yourself to use RAG on AWS." The Managed Knowledge Base type has Amazon Bedrock manage the infrastructure. A Self-managed Knowledge Base is the alternative for teams that need direct control over the vector store.

How it shows up on the exam

Domain D3 / Task 3.1 covers Applications of Foundation Models, including when and how to apply RAG. Questions at this level test whether you can choose the right tool for a scenario, not whether you can configure the pipeline.

Cognitive target: You are expected to recognize that RAG addresses the knowledge gap (information the model was not trained on, or information that has changed) without requiring retraining, and that Knowledge Bases is the managed AWS service that implements this pattern.

Signal phrases to watch for in a scenario:

• "The model does not have access to internal company documents" — points toward RAG / Knowledge Bases.
• "Responses are based on outdated information" — RAG with updated data sources, not retraining.
• "Include citations so users can verify answers" — a documented capability of Knowledge Bases.
• "Reduce the cost of keeping the model current" — RAG is described as cost-effective because it removes the need to continually retrain the model to use private data.

Candidates often confuse grounding (RAG: providing factual context at inference time) with fine-tuning (adjusting model weights for style or specialized reasoning). When a scenario emphasizes what the model knows, think RAG. When it emphasizes how the model behaves, think fine-tuning.

Related concepts

• Retrieval-Augmented Generation — the underlying technique that Knowledge Bases implements
• RAG design considerations — chunking strategy, embedding model choice, retrieval tuning
• AI agents — Knowledge Bases integrates natively with Amazon Bedrock Agents for multi-step, agentic workflows