About event

Single-vector RAG works well for basic retrieval scenarios but quickly reaches its limits when high-precision retrieval or search across complex documents, tables, charts, and visual content is required.

Join us for a practical SPD Talk live stream where we will explore modern Information Retrieval approaches based on Multi-Vector models and the Late Interaction architecture. Through real-world examples, we will demonstrate how ColBERT and ColPali work, how they differ from traditional embedding-based solutions, and how they can be used to build more accurate AI systems.

What we’ll cover:

• Compare Single-Vector and Multi-Vector retrieval approaches;
• Explore the Late Interaction mechanism and its advantages;
• Build token-level retrieval using ColBERT;
• Learn how to search documents as images with ColPali, without relying on OCR;
• Discuss production-ready optimization techniques, including MUVERA, quantization, token pooling, and multi-stage retrieval;
• Examine the trade-offs between retrieval quality, latency, and infrastructure costs.
  
  At the end of the session, participants will receive a hands-on assignment and a Google Colab notebook containing preconfigured code to experiment with multi-vector retrieval on a benchmark dataset. You’ll be able to build your own retrieval pipeline and explore how different configurations impact retrieval quality.
  
  For whom:
  
  This session is tailored for AI/ML & MLOps Engineers, Tech Leads, and Applied Scientists building advanced RAG or Information Retrieval systems. Join us if you want to move beyond traditional single-vector embeddings and explore production-ready multi-vector architectures (ColBERT, ColPali) at scale.