Remember when search meant writing +shoes +comfortable AND standing -heels?
We have come a long way. But here is the plot twist: we did not throw away Boolean operators, TF IDF scores, or inverted indices. We orchestrated them.

Drawing from experience building search and AI systems for Shutterstock 500M plus asset library spanning images, video, music, and sound effects, this talk explores how modern agentic systems weave together twenty five years of information retrieval breakthroughs. When you serve millions of creative professionals who might describe a feeling, hum a melody, upload a reference image, or speak a complex concept, you quickly learn what works and why classical techniques remain irreplaceable.

What You Will Learn

Classical IR as the foundation
BM25 and lexical search remain the backbone of retrieval even as transformers dominate the conversation. These techniques provide precision and controllability that neural approaches alone cannot replace.

The multimodal evolution
From text only keywords to semantic search to true multimodal retrieval across images, video, audio, and text. Learn how agents orchestrate cross modal retrieval and fuse results across modalities.

Hybrid fusion strategies
Concrete patterns such as Reciprocal Rank Fusion, weighted combinations, and cascade reranking that consistently outperform purely semantic or purely keyword approaches.

MCP and composable search
How Model Context Protocol exposes keyword search, vector retrieval, and knowledge graphs as tools that a large language model can intelligently orchestrate.

Voice and natural interaction
How voice assistants transform the interface from manual query construction to natural conversation, especially critical for non technical and ecommerce users.

Agent orchestration in practice
How agents decide when to use BM25 for exact matches, embeddings for semantic recall, cross encoders for precision, CLIP for visual search, and knowledge graphs for entity relationships.

The Core Insight

For decades, we built specialized search components. Each was brilliant at its job and terrible at everything else. The breakthrough is not replacing them. It is teaching an agent to conduct the orchestra, knowing when users need phrase match precision, semantic recall, structured knowledge, or visual similarity.

This dramatically reduces friction. No careful query formulation. No modality constraints. No artificial separation between searching words, images, or sounds.

Practical Takeaways

Search Engineers
Patterns for exposing existing infrastructure as agent accessible tools and for measuring quality in non deterministic systems.

AI Engineers
Why retrieval augmented generation pipelines need lexical anchors, how to build multi hop reasoning across heterogeneous backends, and when transformers underperform decades old inverted indices.

Engineering Managers
Clear return on investment cases for hybrid architectures and practical frameworks for evaluating semantic search investments.

The future of search is not replacing the past. It is orchestrating everything we have learned over twenty five years of information retrieval research into systems that let users express their needs naturally through text or voice and reliably retrieve what they are looking for.