Beyond Single Vectors: Multi-Vector Search for Enhanced...

Beyond Single Vectors: Multi-Vector Search for Enhanced Relevance and Interpretability - Praveen Mohan Prasad & Gene Alpert, Amazon Web Services Vector search has significantly improved retrieval relevance by capturing semantic meaning beyond keywords. However, traditional single-vector models often struggle with longer texts and lack transparency in scoring. Multi-vector models address these limitations by preserving token-level representations, enabling fine-grained late interaction between query tokens and document tokens. This leads to enhanced ranking quality and improved interpretability through token-wise scoring. Models such as ColBERT (for text) and ColPaLi (extending to visual understanding) exemplify this paradigm. In this session, we will explore the benefits of multi-vector search, compare it with traditional embedding approaches, and demonstrate how to implement vector search using these models in OpenSearch. The session will also include a live demo on text and image retrieval using multi-vector retrieval techniques.