AWS re:Invent 2023 - Simplify generative AI app development with Agents for Amazon Bedrock (AIM353)

Introduction

Hello everyone, welcome to the session. Thank you for joining us today. My name is Harel Pimpal, and I'm a Senior Manager of Product with Amazon Bedrock. We will be discussing Amazon Bedrock and agents in particular.

Thank you once again for making the trip to Mandalay Bay and fighting all the traffic. We really appreciate it. Joining me today is Mark Roy, Principal Machine Learning Architect at AWS and our guest speaker, Sean Swanner, CTO of Athene Holdings. Thank you, Sean, for being here.

Overview of Amazon Bedrock

Before we jump in, I want to provide a quick overview of Bedrock. We've had several exciting announcements over the past couple of days, especially in the realm of generative AI.

With Bedrock, we aim to make it extremely easy for you to build and scale generative AI applications. We provide capabilities across three layers:

1. Choice of Models: We offer a selection of models, including fine-tuning, continued pre-training, and retrieval-augmented generation (RAG). A dedicated API simplifies these RAG capabilities.
2. Customization: Customization capabilities are available, including fine-tuning and pre-training.
3. Integration: This will be the focus of today's discussion, specifically through the use of agents.

Agents for Amazon Bedrock

Agents help you extend Foundation Models (FMs) to perform tasks by invoking APIs and looking up information. The automation challenges today include writing a lot of code to create prompts, integrating with company systems, and invoking APIs through Lambda functions.

We created agents for Amazon Bedrock to solve these issues, enabling you to take natural language instructions and apply Chain of Thought prompting to create a custom, optimized prompt for orchestration. All of this is done securely and privately, providing you with control, visibility, and security.

In addition to making agents generally available this week, we also introduced the prompt editor and the Chain of Thought Trace to enhance control and visibility.

Key Features of Agents

1. Multi-Step Orchestration: Agents can break down tasks into multiple steps and execute them.
2. Simplified Generative AI App Development: It streamlines building and deploying AI applications.
3. Fully Managed Infrastructure: Amazon takes care of the orchestration for you.

Agent Basics

Mark Roy:

Microphone check, it looks like we've got volume here. Thanks, everybody, for coming on day four of the conference. For the next few minutes, I will walk you through the basics of agents, orchestrations, action groups, and potential use cases.

Agents enable you to create instructions, make APIs and knowledge bases available, and use Bedrock under the hood to respond to requests.

Use Cases

• Meeting Assistant: Helps in listing available meetings and action items and sending emails.
• HR Policy Assistant: Can check vacation policies or approve vacation requests by interacting with knowledge bases and APIs.
• Laptop Support Assistant: Uses action groups for support history and guidance.
• Ticket Triage: Routes tickets to the right person based on skills and policies.
• Product Review Helper: Automatically responds to product reviews.
• Tractor Maintenance: Retrieves information from extensive repair manuals.

Orchestration

Orchestration involves creating a plan for the agent based on available actions and knowledge bases, executing those steps, and returning a response.

Example

Consider an insurance claims agent. When asked to "send reminders for all open claims with missing documents," the agent:

1. Gets the list of open claims.
2. Iterates through each claim to identify missing documents.
3. Sends reminders to each policyholder.
4. Provides a detailed reasoning trace to ensure transparency.

Implementation Details

• Use existing microservices and operational data stores.
• Wrap functionalities as Lambda functions.
• Rich API descriptions (OpenAPI Schema) assist agents in understanding the capabilities.
• Multiple deployment options including application integration.

Advanced Features

• Prompt Editor: Allows advanced users to modify underlying prompts for greater control.
• Reasoning Trace: Provides an audit trail of the agent’s logical steps, enhancing transparency.
• Security: IAM roles can deny or grant access to individual actions, ensuring enterprise-level security.

Case Study: Athene Holdings

Sean Swanner:

At Athene Holdings, we implemented agents to handle our fixed and indexed annuities. We leveraged Bedrock for document processing and created a knowledge base for automated Q&A responses. This reduced time spent on manual documentation and improved efficiency. Our plans include expanding to more datasets and possibly using agents for code generation in the future.

Practical Demo

Mark Roy:

We demonstrated several practical applications, including:

1. Insurance Claims Agent: Managing claim reminders.
2. Customer Relationship Management: Automating meeting prep.
3. Tax Documentation: Using Bedrock knowledge bases for quick answers.

Future Prospects

We also explored how agents can be used to create other agents, further simplifying development tasks and enhancing productivity.

Keywords

• Amazon Bedrock
• Generative AI
• Foundation Models
• Agents
• Customization
• Integration
• Orchestration
• Knowledge Bases
• Action Groups
• Lambda Functions

FAQ

Q: What are the three main layers of Amazon Bedrock capabilities? A: The three main layers are choice of models, customization, and integration.

Q: How do agents in Amazon Bedrock assist in automation? A: Agents enable multi-step orchestration, simplifying the building and deploying of generative AI applications through secure, privacy-compliant processes.

Q: What is the Chain of Thought prompting? A: Chain of Thought prompting involves breaking down tasks into multiple steps and executing them in sequence.

Q: How can I ensure that the agents operate securely? A: You can grant or deny access to individual actions and agents using IAM roles, ensuring enterprise-level security.

Q: How can I customize the logic used by agents? A: You can use the prompt editor to modify the underlying prompts, allowing for advanced customization of agent behavior.

Q: How does Athene Holdings use agents? A: Athene Holdings uses agents for document processing and creating automated Q&A systems, significantly reducing manual efforts and improving efficiency.

Q: Can agents be used in existing applications? A: Yes, agents can be integrated into existing applications, allowing you to invoke them via the API.

This completes the transformed content into an article format with all necessary summaries, keywords, and FAQs.