AWS re:Invent 2025 - Advanced agentic RAG Systems: Deep dive with Amazon Bedrock (AIM425)

Advanced Agentic RAG Systems: Deep Dive with Amazon Bedrock

Overview

This presentation provided a deep dive into the technical implementation of an "event agent" - an intelligent assistant capable of retrieving session information from a knowledge base and storing/retrieving conversation history to personalize the user experience. The key components covered were:

1. Knowledge Base Integration
2. Agent Memory and Personalization
3. Secure Production Deployment with Agent Core Runtime
4. Identity Management with Agent Core Identity

Knowledge Base Integration

• The presenters used AWS Bedrock Knowledge Bases to ingest and index session data for the AWS re:Invent 2024 conference
• This involved:
  • Downloading and splitting session data into 583 individual documents
  • Uploading documents to an S3 bucket
  • Creating an S3 vector store and index using 1024-dimensional Titan text embeddings
  • Configuring and creating the Bedrock Knowledge Base, specifying the embedding model and storage details
  • Ingesting the data into the knowledge base, which took ~2 minutes
• The presenters created a custom "knowledge base search tool" that the agent can use to query the knowledge base and retrieve relevant session information using semantic search

Agent Memory and Personalization

• To provide a personalized experience, the agent utilizes both short-term and long-term memory:
  • Short-term memory stores the raw conversation history between the user and agent
  • Long-term memory extracts and consolidates user preferences, interests, and other relevant insights from the conversation history
• The presenters used Agent Core Memory, a managed service, to implement this memory functionality:
  • Short-term memory stores all conversation messages, identified by actor ID and session ID
  • Long-term memory uses configurable "strategies" (e.g. user preferences, semantic summaries) to extract and consolidate insights from the short-term history
  • Memory is organized into namespaces based on configurable metadata (e.g. actor ID, session ID, strategy ID) for better retrieval and access control
• When the agent is initialized, it retrieves the user's long-term preferences from memory to personalize the experience
• As new messages are added, they are stored in short-term memory and automatically propagated to long-term memory

Secure Production Deployment

• To deploy the agent in a production environment, the presenters used AWS Agent Core Runtime:
  • Agent Core Runtime provides a serverless, scalable runtime for running agent workloads with true session isolation
  • It automatically provisions a separate micro-VM for each user session, ensuring complete isolation of compute, memory, and storage
  • The presenters demonstrated how to package the agent code into a container image and deploy it to Agent Core Runtime
• To secure access to the agent, the presenters used AWS Agent Core Identity:
  • Users authenticate through an identity provider (e.g. Amazon Cognito)
  • The agent retrieves the user's identity token and uses Agent Core Identity to validate it before allowing access
  • This ensures only authorized users can interact with the agent, and their identity is propagated to the agent code

Key Takeaways

• Agentic RAG (Retrieval Augmented Generation) systems can be built using a combination of knowledge bases, agent memory, and secure production infrastructure
• Knowledge bases provide a way to ingest and index structured data (e.g. event session information) for semantic retrieval
• Agent memory, with both short-term and long-term components, enables personalization by storing and extracting insights from user conversations
• Secure production deployment with Agent Core Runtime and Agent Core Identity ensures scalability, isolation, and access control for mission-critical agent applications

Technical Details

• AWS Bedrock Knowledge Bases
  • Used to ingest and index 583 documents on AWS re:Invent 2024 sessions
  • Leveraged 1024-dimensional Titan text embeddings for semantic search
• Agent Core Memory
  • Provided short-term storage of raw conversation messages
  • Extracted long-term insights using configurable "strategies" (e.g. user preferences, semantic summaries)
  • Organized memory into namespaces based on metadata like actor ID and session ID
• Agent Core Runtime
  • Serverless, scalable runtime for running agent workloads
  • Provisioned separate micro-VMs for each user session to ensure isolation
• Agent Core Identity
  • Validated user identity tokens from identity providers (e.g. Amazon Cognito)
  • Propagated user identity information to the agent code

Business Impact

• The presented agentic RAG system enables a highly personalized event assistant experience, tailoring session recommendations and information retrieval to each individual user
• By integrating knowledge bases, agent memory, and secure production infrastructure, the solution can be deployed at scale to support large-scale event applications
• The ability to persistently store and retrieve user preferences and conversation history allows the agent to build a deeper understanding of each user over time, further enhancing the personalized experience

Examples

• The event agent was able to retrieve relevant session information from the knowledge base based on the user's query (e.g. "What sessions are there on security and AI?")
• The agent leveraged the user's long-term preferences, extracted from previous conversations, to provide personalized session recommendations (e.g. "Based on your interest in cloud architecture, AI, and security, here are some recommended sessions...")
• The secure production deployment using Agent Core Runtime and Agent Core Identity ensured that each user's data and interactions were isolated and accessible only to authorized parties.