AIF-C01: Bedrock Ecosystem

Bedrock Guardrails Filters

Hate, Violence, Sexual, Insults (safety classification categories, PII masking)

Amazon Bedrock (FMs)

Serverless Foundation Models via API; no infrastructure/GPU management; pay-per-token; managed RAG & Agents; quick integration

Bedrock Fine-Tuning

Supervised customization; updates model weights using labeled prompt-response pairs to specialize behavior or style

Bedrock Knowledge Bases

RAG implementation; connects foundation models to enterprise data sources in S3 / vector databases for accurate context-retrieved answers

Bedrock Agents

Autonomous action execution; orchestrates multi-step tasks by executing API calls / AWS Lambda functions based on user requests

Bedrock Model Evaluation

Performance assessment; evaluates models using automatic metrics (ROUGE / BLEU) or human-in-the-loop workflows for quality and safety

Bedrock Continued Pre-Training

Domain adaptation; uses large volumes of unlabeled raw text to teach base foundation models specialized industry jargon before task fine-tuning

Guardrails for Amazon Bedrock

Safety & Compliance Controls; filters harmful content / PII redaction / toxic language / topic blocking across foundation models

Provisioned Throughput

Guaranteed Capacity & Performance; reserves dedicated model processing capacity for consistent high-throughput / low-latency workloads

Bedrock Generative AI Lifecycle

FM Selection to Deployment; Select Base Foundation Model -> Prompt Engineering -> RAG / Knowledge Bases -> Fine-Tuning / Continued Pre-training -> Model Evaluation -> Guardrails Setup -> Provisioned Throughput Deployment