Amazon Bedrock
Managed foundation model service, model access, agents, knowledge bases, guardrails
Amazon Bedrock is a fully managed service for accessing foundation models (FMs) from Amazon and third-party providers via a unified API — without managing infrastructure. It is AWS's primary answer to the generative AI wave, enabling enterprises to build GenAI applications at scale with security and compliance built in.
Key Points
- Foundation Models available: Amazon Titan, Anthropic Claude, Meta LLaMA, Mistral, Cohere, Stability AI
- Single API: switch between models without changing code; pay per token
- No data used for training: your data and prompts are private and not shared with model providers
- Amazon Titan: AWS's own family — Titan Text (G1-Express, G1-Lite), Titan Embeddings, Titan Image Generator
- Bedrock Agents: chain LLM calls with tools (Lambda functions) to complete multi-step tasks autonomously
- Bedrock Knowledge Bases: managed RAG — connect to S3, sync docs, automatic chunking/embedding/vector storage
- Bedrock Guardrails: content filtering for harmful content, PII redaction, topic denial, grounding checks
- Bedrock Model Evaluation: compare model responses against ground truth using automated metrics
- Fine-tuning on Bedrock: customise Claude/Titan with your own labelled data (continued pre-training)
- Model Distillation: compress a large teacher model into a smaller, faster student model
- Provisioned Throughput: reserve model capacity for consistent latency in production
| Bedrock Feature | Use Case | Key Detail |
|---|---|---|
| On-demand FM access | Prototyping, chatbots, summarisation | Pay per token, no provisioning |
| Bedrock Agents | Multi-step task automation | Calls APIs, queries DBs, executes code |
| Knowledge Bases (RAG) | Q&A over documents | Auto chunking, OpenSearch/Pinecone backend |
| Guardrails | Safe AI for production | Content filtering, PII, hallucination grounding |
| Fine-tuning | Custom behaviour/tone | JSONL training data, supervised fine-tuning |
| Model Evaluation | Choose the right model | Human and automated evaluation flows |
Real-World Example
Banque Cantonale Vaudoise (BCV) uses Amazon Bedrock with Claude to analyse financial documents, extract key data, and answer compliance questions — with Guardrails ensuring no sensitive PII is returned to end users.