OpenSearchCon Europe 2026

Dense vector retrieval based on single embeddings has become the default approach for semantic search, but it introduces well-known limitations when dealing with long documents, fine-grained relevance signals, or complex queries. Late interaction retrieval models address these issues by allowing queries and documents to interact at a more granular level, typically across token-level or sub-vector representations, while remaining more scalable than full cross-encoders.

This session explores late interaction models from a systems and search-engine perspective, focusing on what it would take to support them efficiently in OpenSearch-based architectures. It discusses indexing and storage strategies for multi-vector representations, candidate generation and pruning techniques, and the role of late interaction scoring as an intermediate stage between first-pass retrieval and full reranking. The talk emphasizes practical trade-offs in latency, memory, and relevance, and outlines how late interaction models could complement existing lexical, neural sparse, and dense vector approaches in production OpenSearch deployments.