Amazon Bedrock Architecture & Implementation

Amazon Bedrock is AWS’s fully managed, serverless foundation model (FM) service. It provides access to leading generative AI models via a single unified API without managing GPU infrastructure, while ensuring customer training data is never used to train base models.

graph TD
    UserApp[Application / Microservice] -->|Boto3 Converse API| Bedrock[Amazon Bedrock Service]

    subgraph Foundation Models
        Bedrock --> Claude[Anthropic Claude 3.5 Sonnet / Haiku]
        Bedrock --> Llama[Meta Llama 3 / 3.1]
        Bedrock --> Titan[Amazon Titan Text & Embeddings]
        Bedrock --> Mistral[Mistral Large]
    end

    subgraph Bedrock Managed Capabilities
        Bedrock --> KB[Knowledge Bases<br/>Managed RAG + OpenSearch Serverless]
        Bedrock --> Agents[Bedrock Agents<br/>ReAct Planning + Lambda Action Groups]
        Bedrock --> Guard[Bedrock Guardrails<br/>PII Masking + Prompt Injection Shield]
    end

---

1. Supported Foundation Models & Selection Guide

Model Family	Provider	Strengths	Best Fit
Claude 3.5 Sonnet	Anthropic	Coding, complex reasoning, computer use, large context (200k)	High-complexity agentic workflows, architecture design
Claude 3 Haiku	Anthropic	Blazing speed, extremely low token cost	High-throughput classification, routing, fast summarization
Llama 3 / 3.1	Meta	Open-weight foundation, cost-effective	General generation, enterprise task completion
Titan Text / Embeddings	AWS	Cost-optimized, deep AWS native compliance	Document embeddings for vector search
Mistral Large	Mistral AI	Multilingual reasoning, code synthesis	European compliance, low latency

---

2. Bedrock Knowledge Bases (Managed RAG)

Bedrock Knowledge Bases automates the entire Retrieval-Augmented Generation pipeline:

1. Connects to data sources in Amazon S3.
2. Automatically chunks documents (fixed, semantic, or hierarchical).
3. Converts text into vector embeddings using Amazon Titan or Cohere.
4. Indexes vectors into Amazon OpenSearch Serverless, Pinecone, or Amazon Aurora PostgreSQL (pgvector).
5. Exposes a single RetrieveAndGenerate API endpoint that queries the vector store, formats the context, and generates citations.

---

3. Bedrock Agents: Autonomous Tool Execution

Bedrock Agents implement the ReAct (Reason + Act) pattern natively:

• Uses OpenAPI 3.0 schemas to define tools.
• Invokes AWS Lambda functions (Action Groups) to interact with enterprise databases and APIs.
• Maintains multi-turn conversation memory automatically.

---

4. Bedrock Guardrails: Defense-in-Depth

Enterprise LLM deployments must be protected against toxicity and jailbreaks. Bedrock Guardrails sits between the user and the model:

• Prompt Injection Filter: Detects and blocks direct and indirect prompt injection attacks.
• Sensitive Information Filters (PII): Redacts SSNs, credit cards, or internal employee IDs from model responses using regex or AWS Comprehend.
• Denied Topics: Blocks discussions outside enterprise scope (e.g. competitors, legal advice).
• Hallucination Detection: Evaluates grounding scores against retrieved reference passages.

---

5. Implementation: The Bedrock Converse API (Python)

The Converse API (converse / converse_stream) provides a unified request/response schema across all models on Bedrock:

import boto3
from botocore.exceptions import ClientError
 
client = boto3.client("bedrock-runtime", region_name="us-east-1")
 
MODEL_ID = "anthropic.claude-3-5-sonnet-20240620-v1:0"
 
messages = [
    {
        "role": "user",
        "content": [{"text": "Explain Kubernetes Pod Security Standards in 3 sentences."}]
    }
]
 
system_prompt = [{"text": "You are a senior cloud infrastructure security architect."}]
 
inference_config = {
    "temperature": 0.2,
    "maxTokens": 500,
    "topP": 0.9
}
 
try:
    response = client.converse(
        modelId=MODEL_ID,
        messages=messages,
        system=system_prompt,
        inferenceConfig=inference_config
    )
    output_text = response["output"]["message"]["content"][0]["text"]
    token_usage = response["usage"]
    print("Response:\n", output_text)
    print(f"\nTokens: In={token_usage['inputTokens']}, Out={token_usage['outputTokens']}")
 
except ClientError as err:
    print(f"Bedrock Error: {err.response['Error']['Message']}")

Across the wiki

• Amazon Bedrock — generative AI platforms (AWS)
• Amazon SageMaker — generative AI platforms (AWS)
• AWS AI Services — generative AI platforms (AWS)
• AWS Serverless — generative AI platforms (AWS)