AWS re:Invent 2025 - Build Advanced Search with Vector, Hybrid, and AI Techniques (ANT314)

Enhancing Search Relevance with Hybrid and Agentic Techniques

Transitioning from Lexical to Hybrid Search

• Lexical search relies on exact word matching, which struggles with typos, segmentation issues, and abstract queries
• Semantic search using vector embeddings can capture meaning, but can mix up exact matches with contextual relevance
• Hybrid search combines the precision of lexical search with the semantic understanding of vector search
  • Hybrid search runs lexical and vector queries in parallel, then normalizes and blends the scores
  • This enables adaptive ranking that can better match user intent
  • Experiments show hybrid search can improve top-1 accuracy by up to 12-13% on standard benchmarks

Implementing Hybrid Search with Open Search

• Open Search provides built-in support for hybrid search with parallel lexical and vector queries
• Open Search's neural plugin simplifies connecting to external embedding models and creating vector-based queries
• Customers like Recruit used Open Search's hybrid capabilities to:
  • Launch a hybrid search solution quickly
  • Achieve up to 10% more bookings via search and 90% fewer "no hit" searches
  • Gain the flexibility to separate lexical and vector scores for smarter ranking

Advancing to Agentic Search

• Agentic workloads differ from traditional search in several key ways:
  • Executed by AI agents that perform iterative, multi-step reasoning
  • Require real-time responses to dense, multi-part queries
  • Leverage large language models (LLMs) and external data/tools
• Limitations of LLMs for agentic search:
  • Exploding context window sizes lead to quality and latency issues
  • Lack of efficient context management between short-term and long-term memory
• Open Search's Agentic Search Capabilities:
  1. Query Planning: Decomposes complex natural language queries into optimized DSL queries
  2. AI-Powered Intelligence: Automatically selects the right search techniques (lexical, semantic, hybrid) based on the query
  3. Flexible Customization: Allows integrating external data sources and tools via the MCP protocol
  4. Multimodal Embeddings: Supports unifying embeddings across text, image, and other modalities
  5. Agentic Memory: Provides short-term, long-term, and session-based memory management for stateful reasoning
  6. Specialized Agents: Pre-built agents for common agentic use cases like flow orchestration and conversational AI

Performance and Cost Optimizations

• Open Search has achieved 11x core engine performance improvements and 2.5x vector search performance gains in the latest 3.3 release
• To handle exploding vector search workloads, Open Search offers:
  • Exact k-NN for high accuracy and performance
  • Disk mode for managing large vector datasets across RAM and disk
  • Native integration with S3 vector storage for cost-effective scaling

Key Takeaways

• Hybrid search combining lexical and semantic techniques can significantly improve search relevance
• Open Search provides robust hybrid search capabilities, enabling customers to launch solutions quickly
• Agentic search is an emerging paradigm for AI-powered, iterative reasoning over data and tools
• Open Search is investing heavily in agentic search features like query planning, multimodal embeddings, and memory management
• Open Search also focuses on performance and cost optimizations to handle the scale of modern search workloads