AWS Certified Generative AI Developer Professional - Preparation (Draft)

AWS Foundation Model Decision Matrix

Amazon Bedrock

Choose Amazon Bedrock when the question mentions:

• Faster time to market
• Minimal Ops
• Security and Compliance
• No need to train models
• Enterprise workload

AWS prefers Bedrock because,

• No infrastructure to maintain
• Data does not leave AWS
• IAM-native security
• Guardrails built-in

Use Amazon Bedrock to integrate a Manage Foundation Model with Enterprise Data while maintaining security and compliance.

Amazon Sagemaker

Choose Amazon Sagemaker when you see:

• Fine-tune the model
• Domain specific behaviour
• Custom endpoints
• Need for ML lifecycle control

It helps to fine-tune or custom host the models and provides monitoring capabilities.

RAG vs Fine-tuning

• Reduce hallucination (RAG)
• Frequently changing data (RAG)
• Static domain knowledge (Fine tune)
• Lower cost & faster iteration (RAG)
• No retraining overhead (RAG)

Inference Configuration

• Temperature
• Low - Factual, deterministic
• High - Creative, risky
• Max tokens
• Controls cost
• Prevents runaway outputs
• Prompt caching
• Reduces cost
• Improves latency for repeated queries

Embedding Model Selection

Default selection would be Amzon Titan Embeddings via Bedrock because,

• Fully managed
• IAM-integrated
• No data egress
• First class support with Bedrock Knowledge Bases

Prefer Third-Party embeddings (E.g, cohere) inside Bedrock only if specific embedding capabilities required and still it needs to be managed.

If you require full-control on the model, use Sagemaker to host the open-source embedding models. This might be the case if one must reuse existing embedding model and wants to continue after aws migration.

Rarely use External Embedding models (E.g., OpenAI), only if data can move out of AWS