Foundation model integration, data management, and compliance

Pick a model, call it properly, feed it well

This is the biggest domain, so it is where the exam hides most of its traps. Three questions run through it: which model, how do you call it, and what do you hand it.

Choosing and calling the model

Amazon Bedrock gives you foundation models behind one managed API. The Converse API uses one message format across models, while InvokeModel passes the model's own JSON body, so switching providers means rewriting the body. Prefer Converse unless you need a feature it lacks. Keep the model ID out of your code. Put it in AWS AppConfig or Parameter Store and a model swap becomes a config change instead of a deployment.

Resilience is part of the design. Cross-Region inference profiles spread requests across Regions to absorb bursts and to reach models with limited regional availability. Wrap calls in a Step Functions circuit breaker with a fallback model, so a bad afternoon degrades the answer instead of killing it. Customization goes up a ladder: prompt, then RAG, then fine-tuning or continued pre-training, then your own hosting on SageMaker AI, where parameter-efficient methods such as LoRA keep adapters small and versionable.

Data and retrieval

Models only see what you send. Validate data before it reaches them (AWS Glue Data Quality, Lambda checks, CloudWatch metrics), normalise it, and chunk it. Fixed-size chunks are simple, hierarchical chunks match on small children and return larger parents, and semantic chunks split on meaning.

Embeddings turn chunks into vectors. Titan Text Embeddings V2 lets you pick 256, 512 or 1,024 dimensions, trading storage and speed against fidelity. Vectors live in a store such as OpenSearch Serverless, Aurora PostgreSQL with pgvector, or S3 Vectors, and Bedrock Knowledge Bases will manage the whole ingest and retrieve loop for you. Metadata filters narrow results by attributes like department or date. Hybrid search blends keyword and vector scoring, and a reranker model reorders the top hits. Stale vectors are silent liars, so schedule or trigger ingestion syncs when sources change.

Prompts are artifacts

Treat prompts like code. Bedrock Prompt Management stores versioned, parameterised templates, and Prompt Flows chains prompts with conditions and Lambda steps. Version them, review them, and log who used which.