AWS re:Invent 2025 - Advanced RAG Architectures: From Basic Retrieval to Agentic RAG (NTA403)

Advanced RAG Architectures: From Basic Retrieval to Agentic RAG

Overview

• Presenters: Vive Mittal and Palvi Nargun, AWS Solutions Architects
• Focus: Improving the accuracy of Retrieval Augmented Generation (RAG) architectures
• Audience: Experienced RAG users looking to enhance their systems

Key Challenges with RAG Architectures

• As RAG applications become more complex, standard retrieval and generation techniques may not be sufficient
• Issues can arise with:
  • Large document volumes
  • Interconnected/complex documents
  • Proprietary data and abbreviations
  • Varied user query patterns

Amazon Bedrock Knowledge Bases

• Managed service for end-to-end RAG workflows
• Provides:
  • Embedding models
  • Chunking strategies
  • Vector stores (OpenSearch, S3, etc.)
  • Language models for generation

Advanced RAG Techniques

1. Conditional Branching:

   • Intelligently select which vector store(s) to query based on the user's question
   • Example: Routing product-specific vs. policy-related queries to different data sources

2. Parallel Branching:

   • Retrieve and combine data from multiple sources to provide a comprehensive response
   • Example: Identifying root cause, finding fix instructions, and checking inventory for a manufacturing issue

3. Query Reformulation:

   • Break down complex, multi-part queries into smaller, more manageable sub-queries
   • Retrieve relevant chunks, rank, and synthesize the final response

Self-Corrective Agentic RAG

• Central agent orchestrates the RAG workflow
• Iteratively checks:
  1. Relevance of retrieved chunks to the original query
  2. Quality and completeness of the generated response
• Selects and applies the appropriate technique(s):
  • Basic RAG
  • Query expansion
  • Query decomposition
  • Retrieve document
  • Evaluate response quality

Additional RAG Optimization Techniques

1. Injection Flow Enhancements:

   • Chunking strategies (fixed, semantic, hierarchical)
   • Foundational model parsing for multi-modal content
   • Metadata labeling for targeted retrieval

2. Retrieval Flow Enhancements:

   • Metadata filtering
   • Chunk reranking
   • Hybrid search (semantic + keyword)

Business Impact and Use Cases

• Improved accuracy and relevance of RAG-powered applications
• Enables more complex, enterprise-grade use cases:
  • Intelligent customer support
  • Autonomous decision-making systems
  • Knowledge-intensive business processes

Key Takeaways

• RAG architectures require a multi-faceted approach to achieve high accuracy
• Techniques like conditional/parallel branching, query reformulation, and self-corrective agentic RAG can significantly enhance performance
• Optimizing both the injection and retrieval flows is crucial for overall RAG system improvement
• Advanced RAG techniques enable more sophisticated, enterprise-ready applications that can handle complex queries and data