Wrap Up

​

What we covered

• Bedrock Foundation Models — hosted LLMs on the serverless Bedrock platform (examples: DeepSeek, Anthropic Claude, Sonnet, Amazon Nova Micro/Pro). These are pre-hosted: pick the model that best fits your needs (reasoning, latency, cost) instead of always choosing the largest.
• Tokens and context windows — how inputs and generated outputs consume tokens, and why both must fit within a model’s published context window.
• Prompt engineering — how to make prompts directive, unambiguous, and constrained so the model reliably returns the expected format and content.
• Building with the AWS SDK — calling Bedrock runtime methods (for example, InvokeModel or Converse) and using control-plane operations for guardrails and agents.
• Integrations and serverless architecture patterns — common flows like fetching data from S3 and exposing a model-backed API via API Gateway + Lambda.
• Safety and filtering — application-level checks (deny lists, regex, pattern matching) and Bedrock Guardrails for policy enforcement at scale.
• Retrieval-augmented generation (RAG) — combining vector search over your documents (for example, PDFs in S3) with model prompts via Bedrock Knowledge Bases.
• Agents and AgentCore — enabling models to call tools, orchestrate multi-step workflows, and interact with external APIs and databases.

​

Tokens & context windows — practical reminder

• Input tokens + output tokens must fit within the model’s context window.
• When designing systems, budget tokens for context (documents, system instructions) and expected output length.
• Trim or summarize long context documents before embedding them into prompts or use RAG to include only top-k relevant snippets.

​

Prompt engineering best practices

• Use explicit instructions, not questions.
• Set the model’s role (e.g., “You are a technical writer…”).
• Include audience and context.
• Constrain output format and length (examples: “Return a JSON object with keys: title, summary, actions”).
• Provide examples or a template when consistency is required.

​

Building applications with the AWS SDK

// Conceptual example - adapt to your SDK version and environment
import { BedrockRuntimeClient, InvokeModelCommand } from "@aws-sdk/client-bedrock-runtime";

const client = new BedrockRuntimeClient({ region: "us-west-2" });

const command = new InvokeModelCommand({
  modelId: "anthropic.claude-v1",
  input: "Summarize the following text in 3 bullet points: ...",
  // include other runtime parameters such as maxTokens or temperature
});

const response = await client.send(command);
console.log(response.output);

• AWS SDKs and Tools
• Amazon Bedrock documentation (start here for concrete SDK examples)

​

Integration patterns and architecture

• Common serverless pattern: API Gateway → Lambda → Bedrock runtime.
• Data flow example: user request → Lambda retrieves context from S3 (or a DB) → performs vector search against a Knowledge Base → constructs prompt → calls Bedrock → returns result to client.
• Monitor and log: enable model invocation logging and use CloudWatch for metrics, traces, and alarms.

​

Safety, filtering, and permissions

• Filter both user-supplied content and model-generated output.
• Start with simple application-level filters (deny lists, regex) and scale to Bedrock Guardrails for centralized policy enforcement.
• Apply least-privilege IAM: restrict roles and policies to only the Bedrock models, regions, and resources the application needs.

​

Retrieval-augmented generation (RAG) and Knowledge Bases

• Use vector embeddings and similarity search to locate the most relevant document passages (for example, from PDFs stored in S3).
• Build prompts that include only the highest-relevance snippets to minimize token usage and improve answer accuracy.
• Bedrock Knowledge Bases can accelerate RAG development and reduce the time-to-market vs. implementing a full RAG pipeline from scratch.

​

Agents and AgentCore

• Agents let models orchestrate multi-step workflows by calling tools (APIs, DBs, external services).
• Agent decision-making controls whether to call a tool, which tool to call, and how many times to iterate to achieve the goal.
• Bedrock Agents provide a serverless way to combine knowledge bases and tools; AgentCore enables advanced customization and control.

​

Skills you’ve gained

​

Next steps and recommendations

• Keep building: integrate generative AI into real projects to move beyond deterministic logic and unlock new product capabilities.
• Continue learning: follow Bedrock releases, model updates, and shared best practices.
• Experiment: try AgentCore, build Knowledge Bases, and test Guardrails in staging to validate policies and flows.
• Build a portfolio: develop small Gen AI-enabled applications to solidify skills and demonstrate value to teams or employers.

• Amazon Bedrock documentation
• AWS SDKs and Tools
• AWS CloudWatch documentation
• Best practices for IAM