Domain 1: FM Integration, Data Management, and Compliance (31%)
This domain is the largest weighting of the exam (31%). It focuses on how to select, configure, and feed data into foundation models securely and efficiently.
1.1 Foundation Model Selection
Understanding Model Families
- Claude (Anthropic): Excellent for reasoning and large context windows (up to 200k tokens).
- Llama (Meta): Popular open-source model, good for fine-tuning.
- Mistral: Efficient and high performance for its size.
- Titan (AWS): Optimized for AWS infrastructure, good for summarization and embeddings.
Selection Criteria
- Latency: Use smaller models (e.g., Haiku) for real-time chat.
- Cost: Balance token pricing vs. performance.
- Context Window: Necessary for long document RAG.
1.2 Data Management for RAG
Ingestion Pipeline
- Source: S3 buckets containing PDFs, TXT, or HTML files.
- Chunking: Breaking documents into smaller pieces (Standard, Hierarchical, or Semantic chunking).
- Embeddings: Converting text to vectors using
Titan Text EmbeddingsorCohere Embed. - Vector Store: Storing vectors in Amazon OpenSearch Serverless (OSS).
Chunking Strategies
- Fixed Size: Simple but can break context.
- Overlapping Tokens: Ensures context is preserved between chunks.
- Hierarchical: parent-child relationship for better retrieval.
1.3 Compliance and Security
- Data Privacy: Bedrock does not use your data to train base models by default.
- Encryption: Use AWS KMS keys for data at rest (S3 and Vector stores).
- Network Isolation: Use Interface VPC Endpoints to keep traffic within the AWS backbone.
Flashcards
1 / 3
❓(Click to reveal)
💡